Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- README.md +96 -0
- inference.py +107 -0
- mlx_audio/audio_adapter.safetensors +3 -0
- mlx_audio/audio_encoder.safetensors +3 -0
- mlx_audio/deepstack_mergers.safetensors +3 -0
- mlx_llm/added_tokens.json +32 -0
- mlx_llm/chat_template.jinja +85 -0
- mlx_llm/config.json +77 -0
- mlx_llm/generation_config.json +6 -0
- mlx_llm/merges.txt +0 -0
- mlx_llm/model.safetensors +3 -0
- mlx_llm/model.safetensors.index.json +915 -0
- mlx_llm/special_tokens_map.json +31 -0
- mlx_llm/tokenizer.json +3 -0
- mlx_llm/tokenizer_config.json +271 -0
- mlx_llm/vocab.json +0 -0
- run_moss.py +94 -0
- scripts/__pycache__/moss_audio_encoder_mlx.cpython-312.pyc +0 -0
- scripts/__pycache__/moss_audio_mel_mlx.cpython-312.pyc +0 -0
- scripts/__pycache__/moss_audio_mlx_bridge_v3.cpython-312.pyc +0 -0
- scripts/assets/mel_filters.npz +3 -0
- scripts/moss_audio_encoder_mlx.py +215 -0
- scripts/moss_audio_mel_mlx.py +192 -0
- scripts/moss_audio_mlx_bridge_v3.py +314 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
mlx_llm/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,96 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
base_model: OpenMOSS-Team/MOSS-Audio-8B-Instruct
|
| 4 |
+
tags:
|
| 5 |
+
- mlx
|
| 6 |
+
- audio
|
| 7 |
+
- moss-audio
|
| 8 |
+
- asr
|
| 9 |
+
- int4
|
| 10 |
+
- apple-silicon
|
| 11 |
+
language:
|
| 12 |
+
- en
|
| 13 |
+
- zh
|
| 14 |
+
pipeline_tag: audio-text-to-text
|
| 15 |
+
library_name: mlx
|
| 16 |
+
---
|
| 17 |
+
|
| 18 |
+
# MOSS-Audio-8B-Instruct-MLX (hybrid: INT4 LLM + BF16 audio)
|
| 19 |
+
|
| 20 |
+
An Apple MLX conversion of **MOSS-Audio-8B-Instruct** — the ASR-strongest MOSS-Audio
|
| 21 |
+
checkpoint — for fast, low-memory inference on Apple Silicon. LLM quantized to uniform
|
| 22 |
+
INT4 (group_size 64); audio encoder + adapter + DeepStack kept in BF16.
|
| 23 |
+
|
| 24 |
+
> **Why this exists.** The community had MLX builds only of the *Thinking* variant.
|
| 25 |
+
> But Thinking is not ASR-optimized: under identical INT4 quantization it mis-spells
|
| 26 |
+
> letter-spoken tickers (e.g. "CRWD" → "CWD") and is unstable. **Instruct** transcribes
|
| 27 |
+
> them correctly. This build brings Instruct's transcription quality to MLX speed/memory.
|
| 28 |
+
|
| 29 |
+
中文：這是 **MOSS-Audio-8B-Instruct** 的 Apple MLX 轉換版（LLM uniform INT4 + audio 路徑 BF16）。
|
| 30 |
+
社群原本只有 *Thinking* 變體的 MLX 版，但 Thinking 非 ASR 優化——相同 INT4 量化下會把唸出字母的
|
| 31 |
+
ticker（如 "CRWD"）辨識成 "CWD" 且不穩定。**Instruct** 辨識正確。本版把 Instruct 的轉錄品質
|
| 32 |
+
帶到 MLX 的速度與記憶體。
|
| 33 |
+
|
| 34 |
+
## Measured (Apple M1 Max 32GB, 28s zh+en clip)
|
| 35 |
+
|
| 36 |
+
| Metric | PyTorch Instruct | **This (Instruct-MLX)** | Thinking-MLX |
|
| 37 |
+
|---|:---:|:---:|:---:|
|
| 38 |
+
| Ticker "CRWD" | C R W D ✅ | **C R W D ✅** | CWD ❌ |
|
| 39 |
+
| English term (TradingView) | ✅ | ✅ | ✅(loops) |
|
| 40 |
+
| Numerals | Chinese chars | **Arabic 47%** | Arabic |
|
| 41 |
+
| Speed | 1.8x realtime | **6–9x** | 5–8x |
|
| 42 |
+
| Peak memory | ~17 GB | **7.85 GB** | 7.85 GB |
|
| 43 |
+
| Disk | 18 GB | **5.9 GB** | 5.9 GB |
|
| 44 |
+
|
| 45 |
+
**Key finding.** Ticker-ASR degradation in the Thinking-MLX builds comes from the
|
| 46 |
+
Thinking/Instruct *training difference*, not from INT4 quantization — under the same
|
| 47 |
+
uniform INT4, Instruct keeps the ticker. So uniform 4-bit suffices; no mixed-precision needed.
|
| 48 |
+
|
| 49 |
+
## Usage
|
| 50 |
+
|
| 51 |
+
```bash
|
| 52 |
+
pip install mlx mlx-lm soundfile numpy
|
| 53 |
+
python inference.py --audio your_clip_16k_mono.wav
|
| 54 |
+
```
|
| 55 |
+
|
| 56 |
+
Transcription with per-segment timestamps (a Traditional-Chinese prompt triggers
|
| 57 |
+
zh-Hant output):
|
| 58 |
+
|
| 59 |
+
```bash
|
| 60 |
+
python run_moss.py --model . --audio clip.wav \
|
| 61 |
+
--prompt "請逐句轉錄這段音訊，每句標註開始時間。" --temp 0 --repetition-penalty 1.02
|
| 62 |
+
```
|
| 63 |
+
|
| 64 |
+
- **Audio**: 16 kHz mono. Encoder window is Whisper-style **30 s max** — chunk longer audio.
|
| 65 |
+
- **Decoding**: use **greedy (temp=0)** for ASR fidelity. `temp>0` removes the rare
|
| 66 |
+
tail digit-loop but degrades content (wrong numerals, out-of-order timestamps).
|
| 67 |
+
- **digit-loop**: occasionally the model fails to emit EOS and repeats a digit token
|
| 68 |
+
at the very tail; post-truncate repeated trailing digits. Quantization weakens EOS;
|
| 69 |
+
it is a known, harmless tail artifact for transcription use.
|
| 70 |
+
|
| 71 |
+
## How it was converted
|
| 72 |
+
|
| 73 |
+
Pure metadata-mapped weight conversion (no retraining):
|
| 74 |
+
|
| 75 |
+
1. `stage1_mapping.py` — verify every MLX target key is sourceable from the PyTorch
|
| 76 |
+
checkpoint; discover the conv layout transform `transpose(0,2,3,1)`
|
| 77 |
+
(PyTorch `[out,in,h,w]` → MLX `[out,h,w,in]`).
|
| 78 |
+
2. `stage2_convert.py` — extract `language_model.*` + `lm_head`, quantize to INT4
|
| 79 |
+
(group_size 64) via mlx; extract audio encoder/adapter/DeepStack, apply the conv
|
| 80 |
+
transpose, save BF16. Output mirrors the RumiLabs bridge layout exactly.
|
| 81 |
+
|
| 82 |
+
## Limitations
|
| 83 |
+
|
| 84 |
+
- 30-second audio window (chunk + offset timestamps for longer input).
|
| 85 |
+
- Tail digit-loop under greedy (post-truncate).
|
| 86 |
+
- Homophone errors on domain terms (e.g. 300均 → 三百軍) — fix with a glossary/post-pass.
|
| 87 |
+
|
| 88 |
+
## Credits
|
| 89 |
+
|
| 90 |
+
- Base model: [OpenMOSS-Team/MOSS-Audio](https://github.com/OpenMOSS/MOSS-Audio) (Apache-2.0)
|
| 91 |
+
- MLX bridge (encoder/mel/DeepStack port): [RumiLabs](https://huggingface.co/RumiLabs) Thinking-MLX builds
|
| 92 |
+
- Instruct→MLX conversion: this work
|
| 93 |
+
|
| 94 |
+
## License
|
| 95 |
+
|
| 96 |
+
Apache-2.0 (inherited from base model).
|
inference.py
ADDED
|
@@ -0,0 +1,107 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Standalone MOSS-Audio-{4B,8B}-Thinking MLX inference.
|
| 2 |
+
|
| 3 |
+
Usage:
|
| 4 |
+
python inference.py --audio path/to/clip.wav [--max-tokens 2048]
|
| 5 |
+
|
| 6 |
+
Both 4B INT4 and 8B hybrid bundles work with this script. Audio-path
|
| 7 |
+
dtype is inferred from the saved adapter weights (`scales` key => INT4).
|
| 8 |
+
"""
|
| 9 |
+
from __future__ import annotations
|
| 10 |
+
import argparse, sys, time
|
| 11 |
+
from pathlib import Path
|
| 12 |
+
|
| 13 |
+
HERE = Path(__file__).resolve().parent
|
| 14 |
+
sys.path.insert(0, str(HERE / "scripts"))
|
| 15 |
+
|
| 16 |
+
import librosa
|
| 17 |
+
import mlx.core as mx
|
| 18 |
+
import numpy as np
|
| 19 |
+
from mlx_lm import load as mlx_load
|
| 20 |
+
from mlx_lm.generate import generate_step
|
| 21 |
+
from mlx_lm.sample_utils import make_sampler, make_logits_processors
|
| 22 |
+
|
| 23 |
+
from moss_audio_mlx_bridge_v3 import (
|
| 24 |
+
load_mlx_audio_path,
|
| 25 |
+
build_mel_spectrogram,
|
| 26 |
+
run_mlx_audio_pipeline,
|
| 27 |
+
install_deepstack_hooks,
|
| 28 |
+
)
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
def main():
|
| 32 |
+
p = argparse.ArgumentParser()
|
| 33 |
+
p.add_argument("--audio", required=True, help="Path to input .wav (16 kHz, mono)")
|
| 34 |
+
p.add_argument("--max-tokens", type=int, default=2048)
|
| 35 |
+
p.add_argument("--repetition-penalty", type=float, default=1.02,
|
| 36 |
+
help="1.02 kills decode-loops without over-penalizing descriptions")
|
| 37 |
+
args = p.parse_args()
|
| 38 |
+
|
| 39 |
+
ad_w = mx.load(str(HERE / "mlx_audio/audio_adapter.safetensors"))
|
| 40 |
+
if "down_proj.scales" in ad_w:
|
| 41 |
+
llm_hidden = ad_w["down_proj.scales"].shape[0]
|
| 42 |
+
int4_audio = True
|
| 43 |
+
else:
|
| 44 |
+
llm_hidden = ad_w["down_proj.weight"].shape[0]
|
| 45 |
+
int4_audio = False
|
| 46 |
+
size_tag = "4B" if llm_hidden == 2560 else "8B"
|
| 47 |
+
print(f"[detect] {size_tag} bundle, audio int4={int4_audio}")
|
| 48 |
+
|
| 49 |
+
print(f"[load] LLM from {HERE / 'mlx_llm'}")
|
| 50 |
+
t0 = time.perf_counter()
|
| 51 |
+
mlx_model, mlx_tokenizer = mlx_load(str(HERE / "mlx_llm"))
|
| 52 |
+
print(f"[load] LLM: {time.perf_counter()-t0:.1f}s")
|
| 53 |
+
|
| 54 |
+
t0 = time.perf_counter()
|
| 55 |
+
encoder, adapter, mergers = load_mlx_audio_path(HERE / "mlx_audio", int4=int4_audio)
|
| 56 |
+
print(f"[load] audio path: {time.perf_counter()-t0:.1f}s")
|
| 57 |
+
|
| 58 |
+
y, _ = librosa.load(args.audio, sr=16000, mono=True)
|
| 59 |
+
y = y.astype(np.float32)
|
| 60 |
+
print(f"[audio] {args.audio} ({len(y)/16000:.1f}s)")
|
| 61 |
+
|
| 62 |
+
# Pure-MLX mel + input_ids (no torch).
|
| 63 |
+
mel, lens, input_ids_mx, audio_token_id = build_mel_spectrogram(y, mlx_tokenizer)
|
| 64 |
+
primary, ds_embeds = run_mlx_audio_pipeline(encoder, adapter, mergers, mel, lens)
|
| 65 |
+
primary = primary.astype(mx.bfloat16)
|
| 66 |
+
ds_embeds = [d.astype(mx.bfloat16) for d in ds_embeds]
|
| 67 |
+
mx.eval(primary, *ds_embeds)
|
| 68 |
+
|
| 69 |
+
del encoder, adapter, mergers, mel, lens
|
| 70 |
+
import gc; gc.collect(); mx.clear_cache()
|
| 71 |
+
|
| 72 |
+
audio_mask = input_ids_mx == audio_token_id
|
| 73 |
+
audio_positions = np.where(np.array(audio_mask[0]))[0]
|
| 74 |
+
text_embeds = mlx_model.model.embed_tokens(input_ids_mx)
|
| 75 |
+
text_np = np.array(text_embeds.astype(mx.float32))
|
| 76 |
+
primary_np = np.array(primary.astype(mx.float32))
|
| 77 |
+
text_np[0, audio_positions, :] = primary_np[0, :, :]
|
| 78 |
+
merged = mx.array(text_np).astype(mx.bfloat16)
|
| 79 |
+
|
| 80 |
+
ds_flat = [d[0] for d in ds_embeds]
|
| 81 |
+
install_deepstack_hooks(mlx_model, ds_flat, audio_positions)
|
| 82 |
+
|
| 83 |
+
sampler = make_sampler(temp=1.0, top_p=1.0, top_k=50)
|
| 84 |
+
logits_processors = make_logits_processors(
|
| 85 |
+
repetition_penalty=args.repetition_penalty, repetition_context_size=20
|
| 86 |
+
) if args.repetition_penalty else None
|
| 87 |
+
|
| 88 |
+
gen_kwargs = dict(
|
| 89 |
+
prompt=input_ids_mx[0], model=mlx_model,
|
| 90 |
+
input_embeddings=merged[0], max_tokens=args.max_tokens, sampler=sampler,
|
| 91 |
+
)
|
| 92 |
+
if logits_processors:
|
| 93 |
+
gen_kwargs["logits_processors"] = logits_processors
|
| 94 |
+
|
| 95 |
+
t0 = time.perf_counter()
|
| 96 |
+
generated = []
|
| 97 |
+
for tok, _ in generate_step(**gen_kwargs):
|
| 98 |
+
generated.append(int(tok))
|
| 99 |
+
if tok == mlx_tokenizer.eos_token_id:
|
| 100 |
+
break
|
| 101 |
+
elapsed = time.perf_counter() - t0
|
| 102 |
+
print(f"[gen] {len(generated)} tokens in {elapsed:.2f}s ({len(generated)/elapsed:.1f} t/s)")
|
| 103 |
+
print(f"\n=== OUTPUT ===\n{mlx_tokenizer.decode(generated)}\n=== END ===")
|
| 104 |
+
|
| 105 |
+
|
| 106 |
+
if __name__ == "__main__":
|
| 107 |
+
main()
|
mlx_audio/audio_adapter.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7cad6dec43c8c29b84842bff4ee7f95212e0a61b0a8aab3c7947362d3291d2f4
|
| 3 |
+
size 109052198
|
mlx_audio/audio_encoder.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:aa2e9a39a5291851e0ed473d0ee8da090af33ed15499bed93b3c9dc4b7f6b452
|
| 3 |
+
size 1287299131
|
mlx_audio/deepstack_mergers.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:54454963494a509c9bc4f4e77cfe7a838398b5087fe325166b62870220c998a3
|
| 3 |
+
size 327156578
|
mlx_llm/added_tokens.json
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"</think>": 151668,
|
| 3 |
+
"</tool_call>": 151658,
|
| 4 |
+
"</tool_response>": 151666,
|
| 5 |
+
"<think>": 151667,
|
| 6 |
+
"<tool_call>": 151657,
|
| 7 |
+
"<tool_response>": 151665,
|
| 8 |
+
"<|assistant|>": 151671,
|
| 9 |
+
"<|box_end|>": 151649,
|
| 10 |
+
"<|box_start|>": 151648,
|
| 11 |
+
"<|endoftext|>": 151643,
|
| 12 |
+
"<|eot|>": 151672,
|
| 13 |
+
"<|file_sep|>": 151664,
|
| 14 |
+
"<|fim_middle|>": 151660,
|
| 15 |
+
"<|fim_pad|>": 151662,
|
| 16 |
+
"<|fim_prefix|>": 151659,
|
| 17 |
+
"<|fim_suffix|>": 151661,
|
| 18 |
+
"<|im_end|>": 151645,
|
| 19 |
+
"<|im_start|>": 151644,
|
| 20 |
+
"<|image_pad|>": 151655,
|
| 21 |
+
"<|object_ref_end|>": 151647,
|
| 22 |
+
"<|object_ref_start|>": 151646,
|
| 23 |
+
"<|quad_end|>": 151651,
|
| 24 |
+
"<|quad_start|>": 151650,
|
| 25 |
+
"<|repo_name|>": 151663,
|
| 26 |
+
"<|system|>": 151669,
|
| 27 |
+
"<|user|>": 151670,
|
| 28 |
+
"<|video_pad|>": 151656,
|
| 29 |
+
"<|vision_end|>": 151653,
|
| 30 |
+
"<|vision_pad|>": 151654,
|
| 31 |
+
"<|vision_start|>": 151652
|
| 32 |
+
}
|
mlx_llm/chat_template.jinja
ADDED
|
@@ -0,0 +1,85 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 27 |
+
{{- '<|im_start|>' + message.role + '\n' + message.content + '<|im_end|>' + '\n' }}
|
| 28 |
+
{%- elif message.role == "assistant" %}
|
| 29 |
+
{%- set content = message.content %}
|
| 30 |
+
{%- set reasoning_content = '' %}
|
| 31 |
+
{%- if message.reasoning_content is defined and message.reasoning_content is not none %}
|
| 32 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 33 |
+
{%- else %}
|
| 34 |
+
{%- if '</think>' in message.content %}
|
| 35 |
+
{%- set content = message.content.split('</think>')[-1].lstrip('\n') %}
|
| 36 |
+
{%- set reasoning_content = message.content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 37 |
+
{%- endif %}
|
| 38 |
+
{%- endif %}
|
| 39 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 40 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 41 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 42 |
+
{%- else %}
|
| 43 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 44 |
+
{%- endif %}
|
| 45 |
+
{%- else %}
|
| 46 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 47 |
+
{%- endif %}
|
| 48 |
+
{%- if message.tool_calls %}
|
| 49 |
+
{%- for tool_call in message.tool_calls %}
|
| 50 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 51 |
+
{{- '\n' }}
|
| 52 |
+
{%- endif %}
|
| 53 |
+
{%- if tool_call.function %}
|
| 54 |
+
{%- set tool_call = tool_call.function %}
|
| 55 |
+
{%- endif %}
|
| 56 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 57 |
+
{{- tool_call.name }}
|
| 58 |
+
{{- '", "arguments": ' }}
|
| 59 |
+
{%- if tool_call.arguments is string %}
|
| 60 |
+
{{- tool_call.arguments }}
|
| 61 |
+
{%- else %}
|
| 62 |
+
{{- tool_call.arguments | tojson }}
|
| 63 |
+
{%- endif %}
|
| 64 |
+
{{- '}\n</tool_call>' }}
|
| 65 |
+
{%- endfor %}
|
| 66 |
+
{%- endif %}
|
| 67 |
+
{{- '<|im_end|>\n' }}
|
| 68 |
+
{%- elif message.role == "tool" %}
|
| 69 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 70 |
+
{{- '<|im_start|>user' }}
|
| 71 |
+
{%- endif %}
|
| 72 |
+
{{- '\n<tool_response>\n' }}
|
| 73 |
+
{{- message.content }}
|
| 74 |
+
{{- '\n</tool_response>' }}
|
| 75 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 76 |
+
{{- '<|im_end|>\n' }}
|
| 77 |
+
{%- endif %}
|
| 78 |
+
{%- endif %}
|
| 79 |
+
{%- endfor %}
|
| 80 |
+
{%- if add_generation_prompt %}
|
| 81 |
+
{{- '<|im_start|>assistant\n' }}
|
| 82 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 83 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 84 |
+
{%- endif %}
|
| 85 |
+
{%- endif %}
|
mlx_llm/config.json
ADDED
|
@@ -0,0 +1,77 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3ForCausalLM"
|
| 4 |
+
],
|
| 5 |
+
"attention_bias": false,
|
| 6 |
+
"attention_dropout": 0.0,
|
| 7 |
+
"bos_token_id": 151643,
|
| 8 |
+
"eos_token_id": 151645,
|
| 9 |
+
"head_dim": 128,
|
| 10 |
+
"hidden_act": "silu",
|
| 11 |
+
"hidden_size": 4096,
|
| 12 |
+
"initializer_range": 0.02,
|
| 13 |
+
"intermediate_size": 12288,
|
| 14 |
+
"layer_types": [
|
| 15 |
+
"full_attention",
|
| 16 |
+
"full_attention",
|
| 17 |
+
"full_attention",
|
| 18 |
+
"full_attention",
|
| 19 |
+
"full_attention",
|
| 20 |
+
"full_attention",
|
| 21 |
+
"full_attention",
|
| 22 |
+
"full_attention",
|
| 23 |
+
"full_attention",
|
| 24 |
+
"full_attention",
|
| 25 |
+
"full_attention",
|
| 26 |
+
"full_attention",
|
| 27 |
+
"full_attention",
|
| 28 |
+
"full_attention",
|
| 29 |
+
"full_attention",
|
| 30 |
+
"full_attention",
|
| 31 |
+
"full_attention",
|
| 32 |
+
"full_attention",
|
| 33 |
+
"full_attention",
|
| 34 |
+
"full_attention",
|
| 35 |
+
"full_attention",
|
| 36 |
+
"full_attention",
|
| 37 |
+
"full_attention",
|
| 38 |
+
"full_attention",
|
| 39 |
+
"full_attention",
|
| 40 |
+
"full_attention",
|
| 41 |
+
"full_attention",
|
| 42 |
+
"full_attention",
|
| 43 |
+
"full_attention",
|
| 44 |
+
"full_attention",
|
| 45 |
+
"full_attention",
|
| 46 |
+
"full_attention",
|
| 47 |
+
"full_attention",
|
| 48 |
+
"full_attention",
|
| 49 |
+
"full_attention",
|
| 50 |
+
"full_attention"
|
| 51 |
+
],
|
| 52 |
+
"max_position_embeddings": 40960,
|
| 53 |
+
"max_window_layers": 36,
|
| 54 |
+
"model_type": "qwen3",
|
| 55 |
+
"num_attention_heads": 32,
|
| 56 |
+
"num_hidden_layers": 36,
|
| 57 |
+
"num_key_value_heads": 8,
|
| 58 |
+
"quantization": {
|
| 59 |
+
"group_size": 64,
|
| 60 |
+
"bits": 4,
|
| 61 |
+
"mode": "affine"
|
| 62 |
+
},
|
| 63 |
+
"quantization_config": {
|
| 64 |
+
"group_size": 64,
|
| 65 |
+
"bits": 4,
|
| 66 |
+
"mode": "affine"
|
| 67 |
+
},
|
| 68 |
+
"rms_norm_eps": 1e-06,
|
| 69 |
+
"rope_scaling": null,
|
| 70 |
+
"rope_theta": 1000000,
|
| 71 |
+
"sliding_window": null,
|
| 72 |
+
"tie_word_embeddings": false,
|
| 73 |
+
"torch_dtype": "bfloat16",
|
| 74 |
+
"use_cache": true,
|
| 75 |
+
"use_sliding_window": false,
|
| 76 |
+
"vocab_size": 151936
|
| 77 |
+
}
|
mlx_llm/generation_config.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_from_model_config": true,
|
| 3 |
+
"bos_token_id": 151643,
|
| 4 |
+
"eos_token_id": 151645,
|
| 5 |
+
"transformers_version": "4.57.1"
|
| 6 |
+
}
|
mlx_llm/merges.txt
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
mlx_llm/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3125621c811c0d8ec5ad6a0e871b8a76e37ec08d3bb54342c73fbacb1e682433
|
| 3 |
+
size 4607834992
|
mlx_llm/model.safetensors.index.json
ADDED
|
@@ -0,0 +1,915 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"metadata": {
|
| 3 |
+
"total_size": 4607731712,
|
| 4 |
+
"total_parameters": 8190735360
|
| 5 |
+
},
|
| 6 |
+
"weight_map": {
|
| 7 |
+
"lm_head.biases": "model.safetensors",
|
| 8 |
+
"lm_head.scales": "model.safetensors",
|
| 9 |
+
"lm_head.weight": "model.safetensors",
|
| 10 |
+
"model.embed_tokens.biases": "model.safetensors",
|
| 11 |
+
"model.embed_tokens.scales": "model.safetensors",
|
| 12 |
+
"model.embed_tokens.weight": "model.safetensors",
|
| 13 |
+
"model.layers.0.input_layernorm.weight": "model.safetensors",
|
| 14 |
+
"model.layers.0.mlp.down_proj.biases": "model.safetensors",
|
| 15 |
+
"model.layers.0.mlp.down_proj.scales": "model.safetensors",
|
| 16 |
+
"model.layers.0.mlp.down_proj.weight": "model.safetensors",
|
| 17 |
+
"model.layers.0.mlp.gate_proj.biases": "model.safetensors",
|
| 18 |
+
"model.layers.0.mlp.gate_proj.scales": "model.safetensors",
|
| 19 |
+
"model.layers.0.mlp.gate_proj.weight": "model.safetensors",
|
| 20 |
+
"model.layers.0.mlp.up_proj.biases": "model.safetensors",
|
| 21 |
+
"model.layers.0.mlp.up_proj.scales": "model.safetensors",
|
| 22 |
+
"model.layers.0.mlp.up_proj.weight": "model.safetensors",
|
| 23 |
+
"model.layers.0.post_attention_layernorm.weight": "model.safetensors",
|
| 24 |
+
"model.layers.0.self_attn.k_norm.weight": "model.safetensors",
|
| 25 |
+
"model.layers.0.self_attn.k_proj.biases": "model.safetensors",
|
| 26 |
+
"model.layers.0.self_attn.k_proj.scales": "model.safetensors",
|
| 27 |
+
"model.layers.0.self_attn.k_proj.weight": "model.safetensors",
|
| 28 |
+
"model.layers.0.self_attn.o_proj.biases": "model.safetensors",
|
| 29 |
+
"model.layers.0.self_attn.o_proj.scales": "model.safetensors",
|
| 30 |
+
"model.layers.0.self_attn.o_proj.weight": "model.safetensors",
|
| 31 |
+
"model.layers.0.self_attn.q_norm.weight": "model.safetensors",
|
| 32 |
+
"model.layers.0.self_attn.q_proj.biases": "model.safetensors",
|
| 33 |
+
"model.layers.0.self_attn.q_proj.scales": "model.safetensors",
|
| 34 |
+
"model.layers.0.self_attn.q_proj.weight": "model.safetensors",
|
| 35 |
+
"model.layers.0.self_attn.v_proj.biases": "model.safetensors",
|
| 36 |
+
"model.layers.0.self_attn.v_proj.scales": "model.safetensors",
|
| 37 |
+
"model.layers.0.self_attn.v_proj.weight": "model.safetensors",
|
| 38 |
+
"model.layers.1.input_layernorm.weight": "model.safetensors",
|
| 39 |
+
"model.layers.1.mlp.down_proj.biases": "model.safetensors",
|
| 40 |
+
"model.layers.1.mlp.down_proj.scales": "model.safetensors",
|
| 41 |
+
"model.layers.1.mlp.down_proj.weight": "model.safetensors",
|
| 42 |
+
"model.layers.1.mlp.gate_proj.biases": "model.safetensors",
|
| 43 |
+
"model.layers.1.mlp.gate_proj.scales": "model.safetensors",
|
| 44 |
+
"model.layers.1.mlp.gate_proj.weight": "model.safetensors",
|
| 45 |
+
"model.layers.1.mlp.up_proj.biases": "model.safetensors",
|
| 46 |
+
"model.layers.1.mlp.up_proj.scales": "model.safetensors",
|
| 47 |
+
"model.layers.1.mlp.up_proj.weight": "model.safetensors",
|
| 48 |
+
"model.layers.1.post_attention_layernorm.weight": "model.safetensors",
|
| 49 |
+
"model.layers.1.self_attn.k_norm.weight": "model.safetensors",
|
| 50 |
+
"model.layers.1.self_attn.k_proj.biases": "model.safetensors",
|
| 51 |
+
"model.layers.1.self_attn.k_proj.scales": "model.safetensors",
|
| 52 |
+
"model.layers.1.self_attn.k_proj.weight": "model.safetensors",
|
| 53 |
+
"model.layers.1.self_attn.o_proj.biases": "model.safetensors",
|
| 54 |
+
"model.layers.1.self_attn.o_proj.scales": "model.safetensors",
|
| 55 |
+
"model.layers.1.self_attn.o_proj.weight": "model.safetensors",
|
| 56 |
+
"model.layers.1.self_attn.q_norm.weight": "model.safetensors",
|
| 57 |
+
"model.layers.1.self_attn.q_proj.biases": "model.safetensors",
|
| 58 |
+
"model.layers.1.self_attn.q_proj.scales": "model.safetensors",
|
| 59 |
+
"model.layers.1.self_attn.q_proj.weight": "model.safetensors",
|
| 60 |
+
"model.layers.1.self_attn.v_proj.biases": "model.safetensors",
|
| 61 |
+
"model.layers.1.self_attn.v_proj.scales": "model.safetensors",
|
| 62 |
+
"model.layers.1.self_attn.v_proj.weight": "model.safetensors",
|
| 63 |
+
"model.layers.10.input_layernorm.weight": "model.safetensors",
|
| 64 |
+
"model.layers.10.mlp.down_proj.biases": "model.safetensors",
|
| 65 |
+
"model.layers.10.mlp.down_proj.scales": "model.safetensors",
|
| 66 |
+
"model.layers.10.mlp.down_proj.weight": "model.safetensors",
|
| 67 |
+
"model.layers.10.mlp.gate_proj.biases": "model.safetensors",
|
| 68 |
+
"model.layers.10.mlp.gate_proj.scales": "model.safetensors",
|
| 69 |
+
"model.layers.10.mlp.gate_proj.weight": "model.safetensors",
|
| 70 |
+
"model.layers.10.mlp.up_proj.biases": "model.safetensors",
|
| 71 |
+
"model.layers.10.mlp.up_proj.scales": "model.safetensors",
|
| 72 |
+
"model.layers.10.mlp.up_proj.weight": "model.safetensors",
|
| 73 |
+
"model.layers.10.post_attention_layernorm.weight": "model.safetensors",
|
| 74 |
+
"model.layers.10.self_attn.k_norm.weight": "model.safetensors",
|
| 75 |
+
"model.layers.10.self_attn.k_proj.biases": "model.safetensors",
|
| 76 |
+
"model.layers.10.self_attn.k_proj.scales": "model.safetensors",
|
| 77 |
+
"model.layers.10.self_attn.k_proj.weight": "model.safetensors",
|
| 78 |
+
"model.layers.10.self_attn.o_proj.biases": "model.safetensors",
|
| 79 |
+
"model.layers.10.self_attn.o_proj.scales": "model.safetensors",
|
| 80 |
+
"model.layers.10.self_attn.o_proj.weight": "model.safetensors",
|
| 81 |
+
"model.layers.10.self_attn.q_norm.weight": "model.safetensors",
|
| 82 |
+
"model.layers.10.self_attn.q_proj.biases": "model.safetensors",
|
| 83 |
+
"model.layers.10.self_attn.q_proj.scales": "model.safetensors",
|
| 84 |
+
"model.layers.10.self_attn.q_proj.weight": "model.safetensors",
|
| 85 |
+
"model.layers.10.self_attn.v_proj.biases": "model.safetensors",
|
| 86 |
+
"model.layers.10.self_attn.v_proj.scales": "model.safetensors",
|
| 87 |
+
"model.layers.10.self_attn.v_proj.weight": "model.safetensors",
|
| 88 |
+
"model.layers.11.input_layernorm.weight": "model.safetensors",
|
| 89 |
+
"model.layers.11.mlp.down_proj.biases": "model.safetensors",
|
| 90 |
+
"model.layers.11.mlp.down_proj.scales": "model.safetensors",
|
| 91 |
+
"model.layers.11.mlp.down_proj.weight": "model.safetensors",
|
| 92 |
+
"model.layers.11.mlp.gate_proj.biases": "model.safetensors",
|
| 93 |
+
"model.layers.11.mlp.gate_proj.scales": "model.safetensors",
|
| 94 |
+
"model.layers.11.mlp.gate_proj.weight": "model.safetensors",
|
| 95 |
+
"model.layers.11.mlp.up_proj.biases": "model.safetensors",
|
| 96 |
+
"model.layers.11.mlp.up_proj.scales": "model.safetensors",
|
| 97 |
+
"model.layers.11.mlp.up_proj.weight": "model.safetensors",
|
| 98 |
+
"model.layers.11.post_attention_layernorm.weight": "model.safetensors",
|
| 99 |
+
"model.layers.11.self_attn.k_norm.weight": "model.safetensors",
|
| 100 |
+
"model.layers.11.self_attn.k_proj.biases": "model.safetensors",
|
| 101 |
+
"model.layers.11.self_attn.k_proj.scales": "model.safetensors",
|
| 102 |
+
"model.layers.11.self_attn.k_proj.weight": "model.safetensors",
|
| 103 |
+
"model.layers.11.self_attn.o_proj.biases": "model.safetensors",
|
| 104 |
+
"model.layers.11.self_attn.o_proj.scales": "model.safetensors",
|
| 105 |
+
"model.layers.11.self_attn.o_proj.weight": "model.safetensors",
|
| 106 |
+
"model.layers.11.self_attn.q_norm.weight": "model.safetensors",
|
| 107 |
+
"model.layers.11.self_attn.q_proj.biases": "model.safetensors",
|
| 108 |
+
"model.layers.11.self_attn.q_proj.scales": "model.safetensors",
|
| 109 |
+
"model.layers.11.self_attn.q_proj.weight": "model.safetensors",
|
| 110 |
+
"model.layers.11.self_attn.v_proj.biases": "model.safetensors",
|
| 111 |
+
"model.layers.11.self_attn.v_proj.scales": "model.safetensors",
|
| 112 |
+
"model.layers.11.self_attn.v_proj.weight": "model.safetensors",
|
| 113 |
+
"model.layers.12.input_layernorm.weight": "model.safetensors",
|
| 114 |
+
"model.layers.12.mlp.down_proj.biases": "model.safetensors",
|
| 115 |
+
"model.layers.12.mlp.down_proj.scales": "model.safetensors",
|
| 116 |
+
"model.layers.12.mlp.down_proj.weight": "model.safetensors",
|
| 117 |
+
"model.layers.12.mlp.gate_proj.biases": "model.safetensors",
|
| 118 |
+
"model.layers.12.mlp.gate_proj.scales": "model.safetensors",
|
| 119 |
+
"model.layers.12.mlp.gate_proj.weight": "model.safetensors",
|
| 120 |
+
"model.layers.12.mlp.up_proj.biases": "model.safetensors",
|
| 121 |
+
"model.layers.12.mlp.up_proj.scales": "model.safetensors",
|
| 122 |
+
"model.layers.12.mlp.up_proj.weight": "model.safetensors",
|
| 123 |
+
"model.layers.12.post_attention_layernorm.weight": "model.safetensors",
|
| 124 |
+
"model.layers.12.self_attn.k_norm.weight": "model.safetensors",
|
| 125 |
+
"model.layers.12.self_attn.k_proj.biases": "model.safetensors",
|
| 126 |
+
"model.layers.12.self_attn.k_proj.scales": "model.safetensors",
|
| 127 |
+
"model.layers.12.self_attn.k_proj.weight": "model.safetensors",
|
| 128 |
+
"model.layers.12.self_attn.o_proj.biases": "model.safetensors",
|
| 129 |
+
"model.layers.12.self_attn.o_proj.scales": "model.safetensors",
|
| 130 |
+
"model.layers.12.self_attn.o_proj.weight": "model.safetensors",
|
| 131 |
+
"model.layers.12.self_attn.q_norm.weight": "model.safetensors",
|
| 132 |
+
"model.layers.12.self_attn.q_proj.biases": "model.safetensors",
|
| 133 |
+
"model.layers.12.self_attn.q_proj.scales": "model.safetensors",
|
| 134 |
+
"model.layers.12.self_attn.q_proj.weight": "model.safetensors",
|
| 135 |
+
"model.layers.12.self_attn.v_proj.biases": "model.safetensors",
|
| 136 |
+
"model.layers.12.self_attn.v_proj.scales": "model.safetensors",
|
| 137 |
+
"model.layers.12.self_attn.v_proj.weight": "model.safetensors",
|
| 138 |
+
"model.layers.13.input_layernorm.weight": "model.safetensors",
|
| 139 |
+
"model.layers.13.mlp.down_proj.biases": "model.safetensors",
|
| 140 |
+
"model.layers.13.mlp.down_proj.scales": "model.safetensors",
|
| 141 |
+
"model.layers.13.mlp.down_proj.weight": "model.safetensors",
|
| 142 |
+
"model.layers.13.mlp.gate_proj.biases": "model.safetensors",
|
| 143 |
+
"model.layers.13.mlp.gate_proj.scales": "model.safetensors",
|
| 144 |
+
"model.layers.13.mlp.gate_proj.weight": "model.safetensors",
|
| 145 |
+
"model.layers.13.mlp.up_proj.biases": "model.safetensors",
|
| 146 |
+
"model.layers.13.mlp.up_proj.scales": "model.safetensors",
|
| 147 |
+
"model.layers.13.mlp.up_proj.weight": "model.safetensors",
|
| 148 |
+
"model.layers.13.post_attention_layernorm.weight": "model.safetensors",
|
| 149 |
+
"model.layers.13.self_attn.k_norm.weight": "model.safetensors",
|
| 150 |
+
"model.layers.13.self_attn.k_proj.biases": "model.safetensors",
|
| 151 |
+
"model.layers.13.self_attn.k_proj.scales": "model.safetensors",
|
| 152 |
+
"model.layers.13.self_attn.k_proj.weight": "model.safetensors",
|
| 153 |
+
"model.layers.13.self_attn.o_proj.biases": "model.safetensors",
|
| 154 |
+
"model.layers.13.self_attn.o_proj.scales": "model.safetensors",
|
| 155 |
+
"model.layers.13.self_attn.o_proj.weight": "model.safetensors",
|
| 156 |
+
"model.layers.13.self_attn.q_norm.weight": "model.safetensors",
|
| 157 |
+
"model.layers.13.self_attn.q_proj.biases": "model.safetensors",
|
| 158 |
+
"model.layers.13.self_attn.q_proj.scales": "model.safetensors",
|
| 159 |
+
"model.layers.13.self_attn.q_proj.weight": "model.safetensors",
|
| 160 |
+
"model.layers.13.self_attn.v_proj.biases": "model.safetensors",
|
| 161 |
+
"model.layers.13.self_attn.v_proj.scales": "model.safetensors",
|
| 162 |
+
"model.layers.13.self_attn.v_proj.weight": "model.safetensors",
|
| 163 |
+
"model.layers.14.input_layernorm.weight": "model.safetensors",
|
| 164 |
+
"model.layers.14.mlp.down_proj.biases": "model.safetensors",
|
| 165 |
+
"model.layers.14.mlp.down_proj.scales": "model.safetensors",
|
| 166 |
+
"model.layers.14.mlp.down_proj.weight": "model.safetensors",
|
| 167 |
+
"model.layers.14.mlp.gate_proj.biases": "model.safetensors",
|
| 168 |
+
"model.layers.14.mlp.gate_proj.scales": "model.safetensors",
|
| 169 |
+
"model.layers.14.mlp.gate_proj.weight": "model.safetensors",
|
| 170 |
+
"model.layers.14.mlp.up_proj.biases": "model.safetensors",
|
| 171 |
+
"model.layers.14.mlp.up_proj.scales": "model.safetensors",
|
| 172 |
+
"model.layers.14.mlp.up_proj.weight": "model.safetensors",
|
| 173 |
+
"model.layers.14.post_attention_layernorm.weight": "model.safetensors",
|
| 174 |
+
"model.layers.14.self_attn.k_norm.weight": "model.safetensors",
|
| 175 |
+
"model.layers.14.self_attn.k_proj.biases": "model.safetensors",
|
| 176 |
+
"model.layers.14.self_attn.k_proj.scales": "model.safetensors",
|
| 177 |
+
"model.layers.14.self_attn.k_proj.weight": "model.safetensors",
|
| 178 |
+
"model.layers.14.self_attn.o_proj.biases": "model.safetensors",
|
| 179 |
+
"model.layers.14.self_attn.o_proj.scales": "model.safetensors",
|
| 180 |
+
"model.layers.14.self_attn.o_proj.weight": "model.safetensors",
|
| 181 |
+
"model.layers.14.self_attn.q_norm.weight": "model.safetensors",
|
| 182 |
+
"model.layers.14.self_attn.q_proj.biases": "model.safetensors",
|
| 183 |
+
"model.layers.14.self_attn.q_proj.scales": "model.safetensors",
|
| 184 |
+
"model.layers.14.self_attn.q_proj.weight": "model.safetensors",
|
| 185 |
+
"model.layers.14.self_attn.v_proj.biases": "model.safetensors",
|
| 186 |
+
"model.layers.14.self_attn.v_proj.scales": "model.safetensors",
|
| 187 |
+
"model.layers.14.self_attn.v_proj.weight": "model.safetensors",
|
| 188 |
+
"model.layers.15.input_layernorm.weight": "model.safetensors",
|
| 189 |
+
"model.layers.15.mlp.down_proj.biases": "model.safetensors",
|
| 190 |
+
"model.layers.15.mlp.down_proj.scales": "model.safetensors",
|
| 191 |
+
"model.layers.15.mlp.down_proj.weight": "model.safetensors",
|
| 192 |
+
"model.layers.15.mlp.gate_proj.biases": "model.safetensors",
|
| 193 |
+
"model.layers.15.mlp.gate_proj.scales": "model.safetensors",
|
| 194 |
+
"model.layers.15.mlp.gate_proj.weight": "model.safetensors",
|
| 195 |
+
"model.layers.15.mlp.up_proj.biases": "model.safetensors",
|
| 196 |
+
"model.layers.15.mlp.up_proj.scales": "model.safetensors",
|
| 197 |
+
"model.layers.15.mlp.up_proj.weight": "model.safetensors",
|
| 198 |
+
"model.layers.15.post_attention_layernorm.weight": "model.safetensors",
|
| 199 |
+
"model.layers.15.self_attn.k_norm.weight": "model.safetensors",
|
| 200 |
+
"model.layers.15.self_attn.k_proj.biases": "model.safetensors",
|
| 201 |
+
"model.layers.15.self_attn.k_proj.scales": "model.safetensors",
|
| 202 |
+
"model.layers.15.self_attn.k_proj.weight": "model.safetensors",
|
| 203 |
+
"model.layers.15.self_attn.o_proj.biases": "model.safetensors",
|
| 204 |
+
"model.layers.15.self_attn.o_proj.scales": "model.safetensors",
|
| 205 |
+
"model.layers.15.self_attn.o_proj.weight": "model.safetensors",
|
| 206 |
+
"model.layers.15.self_attn.q_norm.weight": "model.safetensors",
|
| 207 |
+
"model.layers.15.self_attn.q_proj.biases": "model.safetensors",
|
| 208 |
+
"model.layers.15.self_attn.q_proj.scales": "model.safetensors",
|
| 209 |
+
"model.layers.15.self_attn.q_proj.weight": "model.safetensors",
|
| 210 |
+
"model.layers.15.self_attn.v_proj.biases": "model.safetensors",
|
| 211 |
+
"model.layers.15.self_attn.v_proj.scales": "model.safetensors",
|
| 212 |
+
"model.layers.15.self_attn.v_proj.weight": "model.safetensors",
|
| 213 |
+
"model.layers.16.input_layernorm.weight": "model.safetensors",
|
| 214 |
+
"model.layers.16.mlp.down_proj.biases": "model.safetensors",
|
| 215 |
+
"model.layers.16.mlp.down_proj.scales": "model.safetensors",
|
| 216 |
+
"model.layers.16.mlp.down_proj.weight": "model.safetensors",
|
| 217 |
+
"model.layers.16.mlp.gate_proj.biases": "model.safetensors",
|
| 218 |
+
"model.layers.16.mlp.gate_proj.scales": "model.safetensors",
|
| 219 |
+
"model.layers.16.mlp.gate_proj.weight": "model.safetensors",
|
| 220 |
+
"model.layers.16.mlp.up_proj.biases": "model.safetensors",
|
| 221 |
+
"model.layers.16.mlp.up_proj.scales": "model.safetensors",
|
| 222 |
+
"model.layers.16.mlp.up_proj.weight": "model.safetensors",
|
| 223 |
+
"model.layers.16.post_attention_layernorm.weight": "model.safetensors",
|
| 224 |
+
"model.layers.16.self_attn.k_norm.weight": "model.safetensors",
|
| 225 |
+
"model.layers.16.self_attn.k_proj.biases": "model.safetensors",
|
| 226 |
+
"model.layers.16.self_attn.k_proj.scales": "model.safetensors",
|
| 227 |
+
"model.layers.16.self_attn.k_proj.weight": "model.safetensors",
|
| 228 |
+
"model.layers.16.self_attn.o_proj.biases": "model.safetensors",
|
| 229 |
+
"model.layers.16.self_attn.o_proj.scales": "model.safetensors",
|
| 230 |
+
"model.layers.16.self_attn.o_proj.weight": "model.safetensors",
|
| 231 |
+
"model.layers.16.self_attn.q_norm.weight": "model.safetensors",
|
| 232 |
+
"model.layers.16.self_attn.q_proj.biases": "model.safetensors",
|
| 233 |
+
"model.layers.16.self_attn.q_proj.scales": "model.safetensors",
|
| 234 |
+
"model.layers.16.self_attn.q_proj.weight": "model.safetensors",
|
| 235 |
+
"model.layers.16.self_attn.v_proj.biases": "model.safetensors",
|
| 236 |
+
"model.layers.16.self_attn.v_proj.scales": "model.safetensors",
|
| 237 |
+
"model.layers.16.self_attn.v_proj.weight": "model.safetensors",
|
| 238 |
+
"model.layers.17.input_layernorm.weight": "model.safetensors",
|
| 239 |
+
"model.layers.17.mlp.down_proj.biases": "model.safetensors",
|
| 240 |
+
"model.layers.17.mlp.down_proj.scales": "model.safetensors",
|
| 241 |
+
"model.layers.17.mlp.down_proj.weight": "model.safetensors",
|
| 242 |
+
"model.layers.17.mlp.gate_proj.biases": "model.safetensors",
|
| 243 |
+
"model.layers.17.mlp.gate_proj.scales": "model.safetensors",
|
| 244 |
+
"model.layers.17.mlp.gate_proj.weight": "model.safetensors",
|
| 245 |
+
"model.layers.17.mlp.up_proj.biases": "model.safetensors",
|
| 246 |
+
"model.layers.17.mlp.up_proj.scales": "model.safetensors",
|
| 247 |
+
"model.layers.17.mlp.up_proj.weight": "model.safetensors",
|
| 248 |
+
"model.layers.17.post_attention_layernorm.weight": "model.safetensors",
|
| 249 |
+
"model.layers.17.self_attn.k_norm.weight": "model.safetensors",
|
| 250 |
+
"model.layers.17.self_attn.k_proj.biases": "model.safetensors",
|
| 251 |
+
"model.layers.17.self_attn.k_proj.scales": "model.safetensors",
|
| 252 |
+
"model.layers.17.self_attn.k_proj.weight": "model.safetensors",
|
| 253 |
+
"model.layers.17.self_attn.o_proj.biases": "model.safetensors",
|
| 254 |
+
"model.layers.17.self_attn.o_proj.scales": "model.safetensors",
|
| 255 |
+
"model.layers.17.self_attn.o_proj.weight": "model.safetensors",
|
| 256 |
+
"model.layers.17.self_attn.q_norm.weight": "model.safetensors",
|
| 257 |
+
"model.layers.17.self_attn.q_proj.biases": "model.safetensors",
|
| 258 |
+
"model.layers.17.self_attn.q_proj.scales": "model.safetensors",
|
| 259 |
+
"model.layers.17.self_attn.q_proj.weight": "model.safetensors",
|
| 260 |
+
"model.layers.17.self_attn.v_proj.biases": "model.safetensors",
|
| 261 |
+
"model.layers.17.self_attn.v_proj.scales": "model.safetensors",
|
| 262 |
+
"model.layers.17.self_attn.v_proj.weight": "model.safetensors",
|
| 263 |
+
"model.layers.18.input_layernorm.weight": "model.safetensors",
|
| 264 |
+
"model.layers.18.mlp.down_proj.biases": "model.safetensors",
|
| 265 |
+
"model.layers.18.mlp.down_proj.scales": "model.safetensors",
|
| 266 |
+
"model.layers.18.mlp.down_proj.weight": "model.safetensors",
|
| 267 |
+
"model.layers.18.mlp.gate_proj.biases": "model.safetensors",
|
| 268 |
+
"model.layers.18.mlp.gate_proj.scales": "model.safetensors",
|
| 269 |
+
"model.layers.18.mlp.gate_proj.weight": "model.safetensors",
|
| 270 |
+
"model.layers.18.mlp.up_proj.biases": "model.safetensors",
|
| 271 |
+
"model.layers.18.mlp.up_proj.scales": "model.safetensors",
|
| 272 |
+
"model.layers.18.mlp.up_proj.weight": "model.safetensors",
|
| 273 |
+
"model.layers.18.post_attention_layernorm.weight": "model.safetensors",
|
| 274 |
+
"model.layers.18.self_attn.k_norm.weight": "model.safetensors",
|
| 275 |
+
"model.layers.18.self_attn.k_proj.biases": "model.safetensors",
|
| 276 |
+
"model.layers.18.self_attn.k_proj.scales": "model.safetensors",
|
| 277 |
+
"model.layers.18.self_attn.k_proj.weight": "model.safetensors",
|
| 278 |
+
"model.layers.18.self_attn.o_proj.biases": "model.safetensors",
|
| 279 |
+
"model.layers.18.self_attn.o_proj.scales": "model.safetensors",
|
| 280 |
+
"model.layers.18.self_attn.o_proj.weight": "model.safetensors",
|
| 281 |
+
"model.layers.18.self_attn.q_norm.weight": "model.safetensors",
|
| 282 |
+
"model.layers.18.self_attn.q_proj.biases": "model.safetensors",
|
| 283 |
+
"model.layers.18.self_attn.q_proj.scales": "model.safetensors",
|
| 284 |
+
"model.layers.18.self_attn.q_proj.weight": "model.safetensors",
|
| 285 |
+
"model.layers.18.self_attn.v_proj.biases": "model.safetensors",
|
| 286 |
+
"model.layers.18.self_attn.v_proj.scales": "model.safetensors",
|
| 287 |
+
"model.layers.18.self_attn.v_proj.weight": "model.safetensors",
|
| 288 |
+
"model.layers.19.input_layernorm.weight": "model.safetensors",
|
| 289 |
+
"model.layers.19.mlp.down_proj.biases": "model.safetensors",
|
| 290 |
+
"model.layers.19.mlp.down_proj.scales": "model.safetensors",
|
| 291 |
+
"model.layers.19.mlp.down_proj.weight": "model.safetensors",
|
| 292 |
+
"model.layers.19.mlp.gate_proj.biases": "model.safetensors",
|
| 293 |
+
"model.layers.19.mlp.gate_proj.scales": "model.safetensors",
|
| 294 |
+
"model.layers.19.mlp.gate_proj.weight": "model.safetensors",
|
| 295 |
+
"model.layers.19.mlp.up_proj.biases": "model.safetensors",
|
| 296 |
+
"model.layers.19.mlp.up_proj.scales": "model.safetensors",
|
| 297 |
+
"model.layers.19.mlp.up_proj.weight": "model.safetensors",
|
| 298 |
+
"model.layers.19.post_attention_layernorm.weight": "model.safetensors",
|
| 299 |
+
"model.layers.19.self_attn.k_norm.weight": "model.safetensors",
|
| 300 |
+
"model.layers.19.self_attn.k_proj.biases": "model.safetensors",
|
| 301 |
+
"model.layers.19.self_attn.k_proj.scales": "model.safetensors",
|
| 302 |
+
"model.layers.19.self_attn.k_proj.weight": "model.safetensors",
|
| 303 |
+
"model.layers.19.self_attn.o_proj.biases": "model.safetensors",
|
| 304 |
+
"model.layers.19.self_attn.o_proj.scales": "model.safetensors",
|
| 305 |
+
"model.layers.19.self_attn.o_proj.weight": "model.safetensors",
|
| 306 |
+
"model.layers.19.self_attn.q_norm.weight": "model.safetensors",
|
| 307 |
+
"model.layers.19.self_attn.q_proj.biases": "model.safetensors",
|
| 308 |
+
"model.layers.19.self_attn.q_proj.scales": "model.safetensors",
|
| 309 |
+
"model.layers.19.self_attn.q_proj.weight": "model.safetensors",
|
| 310 |
+
"model.layers.19.self_attn.v_proj.biases": "model.safetensors",
|
| 311 |
+
"model.layers.19.self_attn.v_proj.scales": "model.safetensors",
|
| 312 |
+
"model.layers.19.self_attn.v_proj.weight": "model.safetensors",
|
| 313 |
+
"model.layers.2.input_layernorm.weight": "model.safetensors",
|
| 314 |
+
"model.layers.2.mlp.down_proj.biases": "model.safetensors",
|
| 315 |
+
"model.layers.2.mlp.down_proj.scales": "model.safetensors",
|
| 316 |
+
"model.layers.2.mlp.down_proj.weight": "model.safetensors",
|
| 317 |
+
"model.layers.2.mlp.gate_proj.biases": "model.safetensors",
|
| 318 |
+
"model.layers.2.mlp.gate_proj.scales": "model.safetensors",
|
| 319 |
+
"model.layers.2.mlp.gate_proj.weight": "model.safetensors",
|
| 320 |
+
"model.layers.2.mlp.up_proj.biases": "model.safetensors",
|
| 321 |
+
"model.layers.2.mlp.up_proj.scales": "model.safetensors",
|
| 322 |
+
"model.layers.2.mlp.up_proj.weight": "model.safetensors",
|
| 323 |
+
"model.layers.2.post_attention_layernorm.weight": "model.safetensors",
|
| 324 |
+
"model.layers.2.self_attn.k_norm.weight": "model.safetensors",
|
| 325 |
+
"model.layers.2.self_attn.k_proj.biases": "model.safetensors",
|
| 326 |
+
"model.layers.2.self_attn.k_proj.scales": "model.safetensors",
|
| 327 |
+
"model.layers.2.self_attn.k_proj.weight": "model.safetensors",
|
| 328 |
+
"model.layers.2.self_attn.o_proj.biases": "model.safetensors",
|
| 329 |
+
"model.layers.2.self_attn.o_proj.scales": "model.safetensors",
|
| 330 |
+
"model.layers.2.self_attn.o_proj.weight": "model.safetensors",
|
| 331 |
+
"model.layers.2.self_attn.q_norm.weight": "model.safetensors",
|
| 332 |
+
"model.layers.2.self_attn.q_proj.biases": "model.safetensors",
|
| 333 |
+
"model.layers.2.self_attn.q_proj.scales": "model.safetensors",
|
| 334 |
+
"model.layers.2.self_attn.q_proj.weight": "model.safetensors",
|
| 335 |
+
"model.layers.2.self_attn.v_proj.biases": "model.safetensors",
|
| 336 |
+
"model.layers.2.self_attn.v_proj.scales": "model.safetensors",
|
| 337 |
+
"model.layers.2.self_attn.v_proj.weight": "model.safetensors",
|
| 338 |
+
"model.layers.20.input_layernorm.weight": "model.safetensors",
|
| 339 |
+
"model.layers.20.mlp.down_proj.biases": "model.safetensors",
|
| 340 |
+
"model.layers.20.mlp.down_proj.scales": "model.safetensors",
|
| 341 |
+
"model.layers.20.mlp.down_proj.weight": "model.safetensors",
|
| 342 |
+
"model.layers.20.mlp.gate_proj.biases": "model.safetensors",
|
| 343 |
+
"model.layers.20.mlp.gate_proj.scales": "model.safetensors",
|
| 344 |
+
"model.layers.20.mlp.gate_proj.weight": "model.safetensors",
|
| 345 |
+
"model.layers.20.mlp.up_proj.biases": "model.safetensors",
|
| 346 |
+
"model.layers.20.mlp.up_proj.scales": "model.safetensors",
|
| 347 |
+
"model.layers.20.mlp.up_proj.weight": "model.safetensors",
|
| 348 |
+
"model.layers.20.post_attention_layernorm.weight": "model.safetensors",
|
| 349 |
+
"model.layers.20.self_attn.k_norm.weight": "model.safetensors",
|
| 350 |
+
"model.layers.20.self_attn.k_proj.biases": "model.safetensors",
|
| 351 |
+
"model.layers.20.self_attn.k_proj.scales": "model.safetensors",
|
| 352 |
+
"model.layers.20.self_attn.k_proj.weight": "model.safetensors",
|
| 353 |
+
"model.layers.20.self_attn.o_proj.biases": "model.safetensors",
|
| 354 |
+
"model.layers.20.self_attn.o_proj.scales": "model.safetensors",
|
| 355 |
+
"model.layers.20.self_attn.o_proj.weight": "model.safetensors",
|
| 356 |
+
"model.layers.20.self_attn.q_norm.weight": "model.safetensors",
|
| 357 |
+
"model.layers.20.self_attn.q_proj.biases": "model.safetensors",
|
| 358 |
+
"model.layers.20.self_attn.q_proj.scales": "model.safetensors",
|
| 359 |
+
"model.layers.20.self_attn.q_proj.weight": "model.safetensors",
|
| 360 |
+
"model.layers.20.self_attn.v_proj.biases": "model.safetensors",
|
| 361 |
+
"model.layers.20.self_attn.v_proj.scales": "model.safetensors",
|
| 362 |
+
"model.layers.20.self_attn.v_proj.weight": "model.safetensors",
|
| 363 |
+
"model.layers.21.input_layernorm.weight": "model.safetensors",
|
| 364 |
+
"model.layers.21.mlp.down_proj.biases": "model.safetensors",
|
| 365 |
+
"model.layers.21.mlp.down_proj.scales": "model.safetensors",
|
| 366 |
+
"model.layers.21.mlp.down_proj.weight": "model.safetensors",
|
| 367 |
+
"model.layers.21.mlp.gate_proj.biases": "model.safetensors",
|
| 368 |
+
"model.layers.21.mlp.gate_proj.scales": "model.safetensors",
|
| 369 |
+
"model.layers.21.mlp.gate_proj.weight": "model.safetensors",
|
| 370 |
+
"model.layers.21.mlp.up_proj.biases": "model.safetensors",
|
| 371 |
+
"model.layers.21.mlp.up_proj.scales": "model.safetensors",
|
| 372 |
+
"model.layers.21.mlp.up_proj.weight": "model.safetensors",
|
| 373 |
+
"model.layers.21.post_attention_layernorm.weight": "model.safetensors",
|
| 374 |
+
"model.layers.21.self_attn.k_norm.weight": "model.safetensors",
|
| 375 |
+
"model.layers.21.self_attn.k_proj.biases": "model.safetensors",
|
| 376 |
+
"model.layers.21.self_attn.k_proj.scales": "model.safetensors",
|
| 377 |
+
"model.layers.21.self_attn.k_proj.weight": "model.safetensors",
|
| 378 |
+
"model.layers.21.self_attn.o_proj.biases": "model.safetensors",
|
| 379 |
+
"model.layers.21.self_attn.o_proj.scales": "model.safetensors",
|
| 380 |
+
"model.layers.21.self_attn.o_proj.weight": "model.safetensors",
|
| 381 |
+
"model.layers.21.self_attn.q_norm.weight": "model.safetensors",
|
| 382 |
+
"model.layers.21.self_attn.q_proj.biases": "model.safetensors",
|
| 383 |
+
"model.layers.21.self_attn.q_proj.scales": "model.safetensors",
|
| 384 |
+
"model.layers.21.self_attn.q_proj.weight": "model.safetensors",
|
| 385 |
+
"model.layers.21.self_attn.v_proj.biases": "model.safetensors",
|
| 386 |
+
"model.layers.21.self_attn.v_proj.scales": "model.safetensors",
|
| 387 |
+
"model.layers.21.self_attn.v_proj.weight": "model.safetensors",
|
| 388 |
+
"model.layers.22.input_layernorm.weight": "model.safetensors",
|
| 389 |
+
"model.layers.22.mlp.down_proj.biases": "model.safetensors",
|
| 390 |
+
"model.layers.22.mlp.down_proj.scales": "model.safetensors",
|
| 391 |
+
"model.layers.22.mlp.down_proj.weight": "model.safetensors",
|
| 392 |
+
"model.layers.22.mlp.gate_proj.biases": "model.safetensors",
|
| 393 |
+
"model.layers.22.mlp.gate_proj.scales": "model.safetensors",
|
| 394 |
+
"model.layers.22.mlp.gate_proj.weight": "model.safetensors",
|
| 395 |
+
"model.layers.22.mlp.up_proj.biases": "model.safetensors",
|
| 396 |
+
"model.layers.22.mlp.up_proj.scales": "model.safetensors",
|
| 397 |
+
"model.layers.22.mlp.up_proj.weight": "model.safetensors",
|
| 398 |
+
"model.layers.22.post_attention_layernorm.weight": "model.safetensors",
|
| 399 |
+
"model.layers.22.self_attn.k_norm.weight": "model.safetensors",
|
| 400 |
+
"model.layers.22.self_attn.k_proj.biases": "model.safetensors",
|
| 401 |
+
"model.layers.22.self_attn.k_proj.scales": "model.safetensors",
|
| 402 |
+
"model.layers.22.self_attn.k_proj.weight": "model.safetensors",
|
| 403 |
+
"model.layers.22.self_attn.o_proj.biases": "model.safetensors",
|
| 404 |
+
"model.layers.22.self_attn.o_proj.scales": "model.safetensors",
|
| 405 |
+
"model.layers.22.self_attn.o_proj.weight": "model.safetensors",
|
| 406 |
+
"model.layers.22.self_attn.q_norm.weight": "model.safetensors",
|
| 407 |
+
"model.layers.22.self_attn.q_proj.biases": "model.safetensors",
|
| 408 |
+
"model.layers.22.self_attn.q_proj.scales": "model.safetensors",
|
| 409 |
+
"model.layers.22.self_attn.q_proj.weight": "model.safetensors",
|
| 410 |
+
"model.layers.22.self_attn.v_proj.biases": "model.safetensors",
|
| 411 |
+
"model.layers.22.self_attn.v_proj.scales": "model.safetensors",
|
| 412 |
+
"model.layers.22.self_attn.v_proj.weight": "model.safetensors",
|
| 413 |
+
"model.layers.23.input_layernorm.weight": "model.safetensors",
|
| 414 |
+
"model.layers.23.mlp.down_proj.biases": "model.safetensors",
|
| 415 |
+
"model.layers.23.mlp.down_proj.scales": "model.safetensors",
|
| 416 |
+
"model.layers.23.mlp.down_proj.weight": "model.safetensors",
|
| 417 |
+
"model.layers.23.mlp.gate_proj.biases": "model.safetensors",
|
| 418 |
+
"model.layers.23.mlp.gate_proj.scales": "model.safetensors",
|
| 419 |
+
"model.layers.23.mlp.gate_proj.weight": "model.safetensors",
|
| 420 |
+
"model.layers.23.mlp.up_proj.biases": "model.safetensors",
|
| 421 |
+
"model.layers.23.mlp.up_proj.scales": "model.safetensors",
|
| 422 |
+
"model.layers.23.mlp.up_proj.weight": "model.safetensors",
|
| 423 |
+
"model.layers.23.post_attention_layernorm.weight": "model.safetensors",
|
| 424 |
+
"model.layers.23.self_attn.k_norm.weight": "model.safetensors",
|
| 425 |
+
"model.layers.23.self_attn.k_proj.biases": "model.safetensors",
|
| 426 |
+
"model.layers.23.self_attn.k_proj.scales": "model.safetensors",
|
| 427 |
+
"model.layers.23.self_attn.k_proj.weight": "model.safetensors",
|
| 428 |
+
"model.layers.23.self_attn.o_proj.biases": "model.safetensors",
|
| 429 |
+
"model.layers.23.self_attn.o_proj.scales": "model.safetensors",
|
| 430 |
+
"model.layers.23.self_attn.o_proj.weight": "model.safetensors",
|
| 431 |
+
"model.layers.23.self_attn.q_norm.weight": "model.safetensors",
|
| 432 |
+
"model.layers.23.self_attn.q_proj.biases": "model.safetensors",
|
| 433 |
+
"model.layers.23.self_attn.q_proj.scales": "model.safetensors",
|
| 434 |
+
"model.layers.23.self_attn.q_proj.weight": "model.safetensors",
|
| 435 |
+
"model.layers.23.self_attn.v_proj.biases": "model.safetensors",
|
| 436 |
+
"model.layers.23.self_attn.v_proj.scales": "model.safetensors",
|
| 437 |
+
"model.layers.23.self_attn.v_proj.weight": "model.safetensors",
|
| 438 |
+
"model.layers.24.input_layernorm.weight": "model.safetensors",
|
| 439 |
+
"model.layers.24.mlp.down_proj.biases": "model.safetensors",
|
| 440 |
+
"model.layers.24.mlp.down_proj.scales": "model.safetensors",
|
| 441 |
+
"model.layers.24.mlp.down_proj.weight": "model.safetensors",
|
| 442 |
+
"model.layers.24.mlp.gate_proj.biases": "model.safetensors",
|
| 443 |
+
"model.layers.24.mlp.gate_proj.scales": "model.safetensors",
|
| 444 |
+
"model.layers.24.mlp.gate_proj.weight": "model.safetensors",
|
| 445 |
+
"model.layers.24.mlp.up_proj.biases": "model.safetensors",
|
| 446 |
+
"model.layers.24.mlp.up_proj.scales": "model.safetensors",
|
| 447 |
+
"model.layers.24.mlp.up_proj.weight": "model.safetensors",
|
| 448 |
+
"model.layers.24.post_attention_layernorm.weight": "model.safetensors",
|
| 449 |
+
"model.layers.24.self_attn.k_norm.weight": "model.safetensors",
|
| 450 |
+
"model.layers.24.self_attn.k_proj.biases": "model.safetensors",
|
| 451 |
+
"model.layers.24.self_attn.k_proj.scales": "model.safetensors",
|
| 452 |
+
"model.layers.24.self_attn.k_proj.weight": "model.safetensors",
|
| 453 |
+
"model.layers.24.self_attn.o_proj.biases": "model.safetensors",
|
| 454 |
+
"model.layers.24.self_attn.o_proj.scales": "model.safetensors",
|
| 455 |
+
"model.layers.24.self_attn.o_proj.weight": "model.safetensors",
|
| 456 |
+
"model.layers.24.self_attn.q_norm.weight": "model.safetensors",
|
| 457 |
+
"model.layers.24.self_attn.q_proj.biases": "model.safetensors",
|
| 458 |
+
"model.layers.24.self_attn.q_proj.scales": "model.safetensors",
|
| 459 |
+
"model.layers.24.self_attn.q_proj.weight": "model.safetensors",
|
| 460 |
+
"model.layers.24.self_attn.v_proj.biases": "model.safetensors",
|
| 461 |
+
"model.layers.24.self_attn.v_proj.scales": "model.safetensors",
|
| 462 |
+
"model.layers.24.self_attn.v_proj.weight": "model.safetensors",
|
| 463 |
+
"model.layers.25.input_layernorm.weight": "model.safetensors",
|
| 464 |
+
"model.layers.25.mlp.down_proj.biases": "model.safetensors",
|
| 465 |
+
"model.layers.25.mlp.down_proj.scales": "model.safetensors",
|
| 466 |
+
"model.layers.25.mlp.down_proj.weight": "model.safetensors",
|
| 467 |
+
"model.layers.25.mlp.gate_proj.biases": "model.safetensors",
|
| 468 |
+
"model.layers.25.mlp.gate_proj.scales": "model.safetensors",
|
| 469 |
+
"model.layers.25.mlp.gate_proj.weight": "model.safetensors",
|
| 470 |
+
"model.layers.25.mlp.up_proj.biases": "model.safetensors",
|
| 471 |
+
"model.layers.25.mlp.up_proj.scales": "model.safetensors",
|
| 472 |
+
"model.layers.25.mlp.up_proj.weight": "model.safetensors",
|
| 473 |
+
"model.layers.25.post_attention_layernorm.weight": "model.safetensors",
|
| 474 |
+
"model.layers.25.self_attn.k_norm.weight": "model.safetensors",
|
| 475 |
+
"model.layers.25.self_attn.k_proj.biases": "model.safetensors",
|
| 476 |
+
"model.layers.25.self_attn.k_proj.scales": "model.safetensors",
|
| 477 |
+
"model.layers.25.self_attn.k_proj.weight": "model.safetensors",
|
| 478 |
+
"model.layers.25.self_attn.o_proj.biases": "model.safetensors",
|
| 479 |
+
"model.layers.25.self_attn.o_proj.scales": "model.safetensors",
|
| 480 |
+
"model.layers.25.self_attn.o_proj.weight": "model.safetensors",
|
| 481 |
+
"model.layers.25.self_attn.q_norm.weight": "model.safetensors",
|
| 482 |
+
"model.layers.25.self_attn.q_proj.biases": "model.safetensors",
|
| 483 |
+
"model.layers.25.self_attn.q_proj.scales": "model.safetensors",
|
| 484 |
+
"model.layers.25.self_attn.q_proj.weight": "model.safetensors",
|
| 485 |
+
"model.layers.25.self_attn.v_proj.biases": "model.safetensors",
|
| 486 |
+
"model.layers.25.self_attn.v_proj.scales": "model.safetensors",
|
| 487 |
+
"model.layers.25.self_attn.v_proj.weight": "model.safetensors",
|
| 488 |
+
"model.layers.26.input_layernorm.weight": "model.safetensors",
|
| 489 |
+
"model.layers.26.mlp.down_proj.biases": "model.safetensors",
|
| 490 |
+
"model.layers.26.mlp.down_proj.scales": "model.safetensors",
|
| 491 |
+
"model.layers.26.mlp.down_proj.weight": "model.safetensors",
|
| 492 |
+
"model.layers.26.mlp.gate_proj.biases": "model.safetensors",
|
| 493 |
+
"model.layers.26.mlp.gate_proj.scales": "model.safetensors",
|
| 494 |
+
"model.layers.26.mlp.gate_proj.weight": "model.safetensors",
|
| 495 |
+
"model.layers.26.mlp.up_proj.biases": "model.safetensors",
|
| 496 |
+
"model.layers.26.mlp.up_proj.scales": "model.safetensors",
|
| 497 |
+
"model.layers.26.mlp.up_proj.weight": "model.safetensors",
|
| 498 |
+
"model.layers.26.post_attention_layernorm.weight": "model.safetensors",
|
| 499 |
+
"model.layers.26.self_attn.k_norm.weight": "model.safetensors",
|
| 500 |
+
"model.layers.26.self_attn.k_proj.biases": "model.safetensors",
|
| 501 |
+
"model.layers.26.self_attn.k_proj.scales": "model.safetensors",
|
| 502 |
+
"model.layers.26.self_attn.k_proj.weight": "model.safetensors",
|
| 503 |
+
"model.layers.26.self_attn.o_proj.biases": "model.safetensors",
|
| 504 |
+
"model.layers.26.self_attn.o_proj.scales": "model.safetensors",
|
| 505 |
+
"model.layers.26.self_attn.o_proj.weight": "model.safetensors",
|
| 506 |
+
"model.layers.26.self_attn.q_norm.weight": "model.safetensors",
|
| 507 |
+
"model.layers.26.self_attn.q_proj.biases": "model.safetensors",
|
| 508 |
+
"model.layers.26.self_attn.q_proj.scales": "model.safetensors",
|
| 509 |
+
"model.layers.26.self_attn.q_proj.weight": "model.safetensors",
|
| 510 |
+
"model.layers.26.self_attn.v_proj.biases": "model.safetensors",
|
| 511 |
+
"model.layers.26.self_attn.v_proj.scales": "model.safetensors",
|
| 512 |
+
"model.layers.26.self_attn.v_proj.weight": "model.safetensors",
|
| 513 |
+
"model.layers.27.input_layernorm.weight": "model.safetensors",
|
| 514 |
+
"model.layers.27.mlp.down_proj.biases": "model.safetensors",
|
| 515 |
+
"model.layers.27.mlp.down_proj.scales": "model.safetensors",
|
| 516 |
+
"model.layers.27.mlp.down_proj.weight": "model.safetensors",
|
| 517 |
+
"model.layers.27.mlp.gate_proj.biases": "model.safetensors",
|
| 518 |
+
"model.layers.27.mlp.gate_proj.scales": "model.safetensors",
|
| 519 |
+
"model.layers.27.mlp.gate_proj.weight": "model.safetensors",
|
| 520 |
+
"model.layers.27.mlp.up_proj.biases": "model.safetensors",
|
| 521 |
+
"model.layers.27.mlp.up_proj.scales": "model.safetensors",
|
| 522 |
+
"model.layers.27.mlp.up_proj.weight": "model.safetensors",
|
| 523 |
+
"model.layers.27.post_attention_layernorm.weight": "model.safetensors",
|
| 524 |
+
"model.layers.27.self_attn.k_norm.weight": "model.safetensors",
|
| 525 |
+
"model.layers.27.self_attn.k_proj.biases": "model.safetensors",
|
| 526 |
+
"model.layers.27.self_attn.k_proj.scales": "model.safetensors",
|
| 527 |
+
"model.layers.27.self_attn.k_proj.weight": "model.safetensors",
|
| 528 |
+
"model.layers.27.self_attn.o_proj.biases": "model.safetensors",
|
| 529 |
+
"model.layers.27.self_attn.o_proj.scales": "model.safetensors",
|
| 530 |
+
"model.layers.27.self_attn.o_proj.weight": "model.safetensors",
|
| 531 |
+
"model.layers.27.self_attn.q_norm.weight": "model.safetensors",
|
| 532 |
+
"model.layers.27.self_attn.q_proj.biases": "model.safetensors",
|
| 533 |
+
"model.layers.27.self_attn.q_proj.scales": "model.safetensors",
|
| 534 |
+
"model.layers.27.self_attn.q_proj.weight": "model.safetensors",
|
| 535 |
+
"model.layers.27.self_attn.v_proj.biases": "model.safetensors",
|
| 536 |
+
"model.layers.27.self_attn.v_proj.scales": "model.safetensors",
|
| 537 |
+
"model.layers.27.self_attn.v_proj.weight": "model.safetensors",
|
| 538 |
+
"model.layers.28.input_layernorm.weight": "model.safetensors",
|
| 539 |
+
"model.layers.28.mlp.down_proj.biases": "model.safetensors",
|
| 540 |
+
"model.layers.28.mlp.down_proj.scales": "model.safetensors",
|
| 541 |
+
"model.layers.28.mlp.down_proj.weight": "model.safetensors",
|
| 542 |
+
"model.layers.28.mlp.gate_proj.biases": "model.safetensors",
|
| 543 |
+
"model.layers.28.mlp.gate_proj.scales": "model.safetensors",
|
| 544 |
+
"model.layers.28.mlp.gate_proj.weight": "model.safetensors",
|
| 545 |
+
"model.layers.28.mlp.up_proj.biases": "model.safetensors",
|
| 546 |
+
"model.layers.28.mlp.up_proj.scales": "model.safetensors",
|
| 547 |
+
"model.layers.28.mlp.up_proj.weight": "model.safetensors",
|
| 548 |
+
"model.layers.28.post_attention_layernorm.weight": "model.safetensors",
|
| 549 |
+
"model.layers.28.self_attn.k_norm.weight": "model.safetensors",
|
| 550 |
+
"model.layers.28.self_attn.k_proj.biases": "model.safetensors",
|
| 551 |
+
"model.layers.28.self_attn.k_proj.scales": "model.safetensors",
|
| 552 |
+
"model.layers.28.self_attn.k_proj.weight": "model.safetensors",
|
| 553 |
+
"model.layers.28.self_attn.o_proj.biases": "model.safetensors",
|
| 554 |
+
"model.layers.28.self_attn.o_proj.scales": "model.safetensors",
|
| 555 |
+
"model.layers.28.self_attn.o_proj.weight": "model.safetensors",
|
| 556 |
+
"model.layers.28.self_attn.q_norm.weight": "model.safetensors",
|
| 557 |
+
"model.layers.28.self_attn.q_proj.biases": "model.safetensors",
|
| 558 |
+
"model.layers.28.self_attn.q_proj.scales": "model.safetensors",
|
| 559 |
+
"model.layers.28.self_attn.q_proj.weight": "model.safetensors",
|
| 560 |
+
"model.layers.28.self_attn.v_proj.biases": "model.safetensors",
|
| 561 |
+
"model.layers.28.self_attn.v_proj.scales": "model.safetensors",
|
| 562 |
+
"model.layers.28.self_attn.v_proj.weight": "model.safetensors",
|
| 563 |
+
"model.layers.29.input_layernorm.weight": "model.safetensors",
|
| 564 |
+
"model.layers.29.mlp.down_proj.biases": "model.safetensors",
|
| 565 |
+
"model.layers.29.mlp.down_proj.scales": "model.safetensors",
|
| 566 |
+
"model.layers.29.mlp.down_proj.weight": "model.safetensors",
|
| 567 |
+
"model.layers.29.mlp.gate_proj.biases": "model.safetensors",
|
| 568 |
+
"model.layers.29.mlp.gate_proj.scales": "model.safetensors",
|
| 569 |
+
"model.layers.29.mlp.gate_proj.weight": "model.safetensors",
|
| 570 |
+
"model.layers.29.mlp.up_proj.biases": "model.safetensors",
|
| 571 |
+
"model.layers.29.mlp.up_proj.scales": "model.safetensors",
|
| 572 |
+
"model.layers.29.mlp.up_proj.weight": "model.safetensors",
|
| 573 |
+
"model.layers.29.post_attention_layernorm.weight": "model.safetensors",
|
| 574 |
+
"model.layers.29.self_attn.k_norm.weight": "model.safetensors",
|
| 575 |
+
"model.layers.29.self_attn.k_proj.biases": "model.safetensors",
|
| 576 |
+
"model.layers.29.self_attn.k_proj.scales": "model.safetensors",
|
| 577 |
+
"model.layers.29.self_attn.k_proj.weight": "model.safetensors",
|
| 578 |
+
"model.layers.29.self_attn.o_proj.biases": "model.safetensors",
|
| 579 |
+
"model.layers.29.self_attn.o_proj.scales": "model.safetensors",
|
| 580 |
+
"model.layers.29.self_attn.o_proj.weight": "model.safetensors",
|
| 581 |
+
"model.layers.29.self_attn.q_norm.weight": "model.safetensors",
|
| 582 |
+
"model.layers.29.self_attn.q_proj.biases": "model.safetensors",
|
| 583 |
+
"model.layers.29.self_attn.q_proj.scales": "model.safetensors",
|
| 584 |
+
"model.layers.29.self_attn.q_proj.weight": "model.safetensors",
|
| 585 |
+
"model.layers.29.self_attn.v_proj.biases": "model.safetensors",
|
| 586 |
+
"model.layers.29.self_attn.v_proj.scales": "model.safetensors",
|
| 587 |
+
"model.layers.29.self_attn.v_proj.weight": "model.safetensors",
|
| 588 |
+
"model.layers.3.input_layernorm.weight": "model.safetensors",
|
| 589 |
+
"model.layers.3.mlp.down_proj.biases": "model.safetensors",
|
| 590 |
+
"model.layers.3.mlp.down_proj.scales": "model.safetensors",
|
| 591 |
+
"model.layers.3.mlp.down_proj.weight": "model.safetensors",
|
| 592 |
+
"model.layers.3.mlp.gate_proj.biases": "model.safetensors",
|
| 593 |
+
"model.layers.3.mlp.gate_proj.scales": "model.safetensors",
|
| 594 |
+
"model.layers.3.mlp.gate_proj.weight": "model.safetensors",
|
| 595 |
+
"model.layers.3.mlp.up_proj.biases": "model.safetensors",
|
| 596 |
+
"model.layers.3.mlp.up_proj.scales": "model.safetensors",
|
| 597 |
+
"model.layers.3.mlp.up_proj.weight": "model.safetensors",
|
| 598 |
+
"model.layers.3.post_attention_layernorm.weight": "model.safetensors",
|
| 599 |
+
"model.layers.3.self_attn.k_norm.weight": "model.safetensors",
|
| 600 |
+
"model.layers.3.self_attn.k_proj.biases": "model.safetensors",
|
| 601 |
+
"model.layers.3.self_attn.k_proj.scales": "model.safetensors",
|
| 602 |
+
"model.layers.3.self_attn.k_proj.weight": "model.safetensors",
|
| 603 |
+
"model.layers.3.self_attn.o_proj.biases": "model.safetensors",
|
| 604 |
+
"model.layers.3.self_attn.o_proj.scales": "model.safetensors",
|
| 605 |
+
"model.layers.3.self_attn.o_proj.weight": "model.safetensors",
|
| 606 |
+
"model.layers.3.self_attn.q_norm.weight": "model.safetensors",
|
| 607 |
+
"model.layers.3.self_attn.q_proj.biases": "model.safetensors",
|
| 608 |
+
"model.layers.3.self_attn.q_proj.scales": "model.safetensors",
|
| 609 |
+
"model.layers.3.self_attn.q_proj.weight": "model.safetensors",
|
| 610 |
+
"model.layers.3.self_attn.v_proj.biases": "model.safetensors",
|
| 611 |
+
"model.layers.3.self_attn.v_proj.scales": "model.safetensors",
|
| 612 |
+
"model.layers.3.self_attn.v_proj.weight": "model.safetensors",
|
| 613 |
+
"model.layers.30.input_layernorm.weight": "model.safetensors",
|
| 614 |
+
"model.layers.30.mlp.down_proj.biases": "model.safetensors",
|
| 615 |
+
"model.layers.30.mlp.down_proj.scales": "model.safetensors",
|
| 616 |
+
"model.layers.30.mlp.down_proj.weight": "model.safetensors",
|
| 617 |
+
"model.layers.30.mlp.gate_proj.biases": "model.safetensors",
|
| 618 |
+
"model.layers.30.mlp.gate_proj.scales": "model.safetensors",
|
| 619 |
+
"model.layers.30.mlp.gate_proj.weight": "model.safetensors",
|
| 620 |
+
"model.layers.30.mlp.up_proj.biases": "model.safetensors",
|
| 621 |
+
"model.layers.30.mlp.up_proj.scales": "model.safetensors",
|
| 622 |
+
"model.layers.30.mlp.up_proj.weight": "model.safetensors",
|
| 623 |
+
"model.layers.30.post_attention_layernorm.weight": "model.safetensors",
|
| 624 |
+
"model.layers.30.self_attn.k_norm.weight": "model.safetensors",
|
| 625 |
+
"model.layers.30.self_attn.k_proj.biases": "model.safetensors",
|
| 626 |
+
"model.layers.30.self_attn.k_proj.scales": "model.safetensors",
|
| 627 |
+
"model.layers.30.self_attn.k_proj.weight": "model.safetensors",
|
| 628 |
+
"model.layers.30.self_attn.o_proj.biases": "model.safetensors",
|
| 629 |
+
"model.layers.30.self_attn.o_proj.scales": "model.safetensors",
|
| 630 |
+
"model.layers.30.self_attn.o_proj.weight": "model.safetensors",
|
| 631 |
+
"model.layers.30.self_attn.q_norm.weight": "model.safetensors",
|
| 632 |
+
"model.layers.30.self_attn.q_proj.biases": "model.safetensors",
|
| 633 |
+
"model.layers.30.self_attn.q_proj.scales": "model.safetensors",
|
| 634 |
+
"model.layers.30.self_attn.q_proj.weight": "model.safetensors",
|
| 635 |
+
"model.layers.30.self_attn.v_proj.biases": "model.safetensors",
|
| 636 |
+
"model.layers.30.self_attn.v_proj.scales": "model.safetensors",
|
| 637 |
+
"model.layers.30.self_attn.v_proj.weight": "model.safetensors",
|
| 638 |
+
"model.layers.31.input_layernorm.weight": "model.safetensors",
|
| 639 |
+
"model.layers.31.mlp.down_proj.biases": "model.safetensors",
|
| 640 |
+
"model.layers.31.mlp.down_proj.scales": "model.safetensors",
|
| 641 |
+
"model.layers.31.mlp.down_proj.weight": "model.safetensors",
|
| 642 |
+
"model.layers.31.mlp.gate_proj.biases": "model.safetensors",
|
| 643 |
+
"model.layers.31.mlp.gate_proj.scales": "model.safetensors",
|
| 644 |
+
"model.layers.31.mlp.gate_proj.weight": "model.safetensors",
|
| 645 |
+
"model.layers.31.mlp.up_proj.biases": "model.safetensors",
|
| 646 |
+
"model.layers.31.mlp.up_proj.scales": "model.safetensors",
|
| 647 |
+
"model.layers.31.mlp.up_proj.weight": "model.safetensors",
|
| 648 |
+
"model.layers.31.post_attention_layernorm.weight": "model.safetensors",
|
| 649 |
+
"model.layers.31.self_attn.k_norm.weight": "model.safetensors",
|
| 650 |
+
"model.layers.31.self_attn.k_proj.biases": "model.safetensors",
|
| 651 |
+
"model.layers.31.self_attn.k_proj.scales": "model.safetensors",
|
| 652 |
+
"model.layers.31.self_attn.k_proj.weight": "model.safetensors",
|
| 653 |
+
"model.layers.31.self_attn.o_proj.biases": "model.safetensors",
|
| 654 |
+
"model.layers.31.self_attn.o_proj.scales": "model.safetensors",
|
| 655 |
+
"model.layers.31.self_attn.o_proj.weight": "model.safetensors",
|
| 656 |
+
"model.layers.31.self_attn.q_norm.weight": "model.safetensors",
|
| 657 |
+
"model.layers.31.self_attn.q_proj.biases": "model.safetensors",
|
| 658 |
+
"model.layers.31.self_attn.q_proj.scales": "model.safetensors",
|
| 659 |
+
"model.layers.31.self_attn.q_proj.weight": "model.safetensors",
|
| 660 |
+
"model.layers.31.self_attn.v_proj.biases": "model.safetensors",
|
| 661 |
+
"model.layers.31.self_attn.v_proj.scales": "model.safetensors",
|
| 662 |
+
"model.layers.31.self_attn.v_proj.weight": "model.safetensors",
|
| 663 |
+
"model.layers.32.input_layernorm.weight": "model.safetensors",
|
| 664 |
+
"model.layers.32.mlp.down_proj.biases": "model.safetensors",
|
| 665 |
+
"model.layers.32.mlp.down_proj.scales": "model.safetensors",
|
| 666 |
+
"model.layers.32.mlp.down_proj.weight": "model.safetensors",
|
| 667 |
+
"model.layers.32.mlp.gate_proj.biases": "model.safetensors",
|
| 668 |
+
"model.layers.32.mlp.gate_proj.scales": "model.safetensors",
|
| 669 |
+
"model.layers.32.mlp.gate_proj.weight": "model.safetensors",
|
| 670 |
+
"model.layers.32.mlp.up_proj.biases": "model.safetensors",
|
| 671 |
+
"model.layers.32.mlp.up_proj.scales": "model.safetensors",
|
| 672 |
+
"model.layers.32.mlp.up_proj.weight": "model.safetensors",
|
| 673 |
+
"model.layers.32.post_attention_layernorm.weight": "model.safetensors",
|
| 674 |
+
"model.layers.32.self_attn.k_norm.weight": "model.safetensors",
|
| 675 |
+
"model.layers.32.self_attn.k_proj.biases": "model.safetensors",
|
| 676 |
+
"model.layers.32.self_attn.k_proj.scales": "model.safetensors",
|
| 677 |
+
"model.layers.32.self_attn.k_proj.weight": "model.safetensors",
|
| 678 |
+
"model.layers.32.self_attn.o_proj.biases": "model.safetensors",
|
| 679 |
+
"model.layers.32.self_attn.o_proj.scales": "model.safetensors",
|
| 680 |
+
"model.layers.32.self_attn.o_proj.weight": "model.safetensors",
|
| 681 |
+
"model.layers.32.self_attn.q_norm.weight": "model.safetensors",
|
| 682 |
+
"model.layers.32.self_attn.q_proj.biases": "model.safetensors",
|
| 683 |
+
"model.layers.32.self_attn.q_proj.scales": "model.safetensors",
|
| 684 |
+
"model.layers.32.self_attn.q_proj.weight": "model.safetensors",
|
| 685 |
+
"model.layers.32.self_attn.v_proj.biases": "model.safetensors",
|
| 686 |
+
"model.layers.32.self_attn.v_proj.scales": "model.safetensors",
|
| 687 |
+
"model.layers.32.self_attn.v_proj.weight": "model.safetensors",
|
| 688 |
+
"model.layers.33.input_layernorm.weight": "model.safetensors",
|
| 689 |
+
"model.layers.33.mlp.down_proj.biases": "model.safetensors",
|
| 690 |
+
"model.layers.33.mlp.down_proj.scales": "model.safetensors",
|
| 691 |
+
"model.layers.33.mlp.down_proj.weight": "model.safetensors",
|
| 692 |
+
"model.layers.33.mlp.gate_proj.biases": "model.safetensors",
|
| 693 |
+
"model.layers.33.mlp.gate_proj.scales": "model.safetensors",
|
| 694 |
+
"model.layers.33.mlp.gate_proj.weight": "model.safetensors",
|
| 695 |
+
"model.layers.33.mlp.up_proj.biases": "model.safetensors",
|
| 696 |
+
"model.layers.33.mlp.up_proj.scales": "model.safetensors",
|
| 697 |
+
"model.layers.33.mlp.up_proj.weight": "model.safetensors",
|
| 698 |
+
"model.layers.33.post_attention_layernorm.weight": "model.safetensors",
|
| 699 |
+
"model.layers.33.self_attn.k_norm.weight": "model.safetensors",
|
| 700 |
+
"model.layers.33.self_attn.k_proj.biases": "model.safetensors",
|
| 701 |
+
"model.layers.33.self_attn.k_proj.scales": "model.safetensors",
|
| 702 |
+
"model.layers.33.self_attn.k_proj.weight": "model.safetensors",
|
| 703 |
+
"model.layers.33.self_attn.o_proj.biases": "model.safetensors",
|
| 704 |
+
"model.layers.33.self_attn.o_proj.scales": "model.safetensors",
|
| 705 |
+
"model.layers.33.self_attn.o_proj.weight": "model.safetensors",
|
| 706 |
+
"model.layers.33.self_attn.q_norm.weight": "model.safetensors",
|
| 707 |
+
"model.layers.33.self_attn.q_proj.biases": "model.safetensors",
|
| 708 |
+
"model.layers.33.self_attn.q_proj.scales": "model.safetensors",
|
| 709 |
+
"model.layers.33.self_attn.q_proj.weight": "model.safetensors",
|
| 710 |
+
"model.layers.33.self_attn.v_proj.biases": "model.safetensors",
|
| 711 |
+
"model.layers.33.self_attn.v_proj.scales": "model.safetensors",
|
| 712 |
+
"model.layers.33.self_attn.v_proj.weight": "model.safetensors",
|
| 713 |
+
"model.layers.34.input_layernorm.weight": "model.safetensors",
|
| 714 |
+
"model.layers.34.mlp.down_proj.biases": "model.safetensors",
|
| 715 |
+
"model.layers.34.mlp.down_proj.scales": "model.safetensors",
|
| 716 |
+
"model.layers.34.mlp.down_proj.weight": "model.safetensors",
|
| 717 |
+
"model.layers.34.mlp.gate_proj.biases": "model.safetensors",
|
| 718 |
+
"model.layers.34.mlp.gate_proj.scales": "model.safetensors",
|
| 719 |
+
"model.layers.34.mlp.gate_proj.weight": "model.safetensors",
|
| 720 |
+
"model.layers.34.mlp.up_proj.biases": "model.safetensors",
|
| 721 |
+
"model.layers.34.mlp.up_proj.scales": "model.safetensors",
|
| 722 |
+
"model.layers.34.mlp.up_proj.weight": "model.safetensors",
|
| 723 |
+
"model.layers.34.post_attention_layernorm.weight": "model.safetensors",
|
| 724 |
+
"model.layers.34.self_attn.k_norm.weight": "model.safetensors",
|
| 725 |
+
"model.layers.34.self_attn.k_proj.biases": "model.safetensors",
|
| 726 |
+
"model.layers.34.self_attn.k_proj.scales": "model.safetensors",
|
| 727 |
+
"model.layers.34.self_attn.k_proj.weight": "model.safetensors",
|
| 728 |
+
"model.layers.34.self_attn.o_proj.biases": "model.safetensors",
|
| 729 |
+
"model.layers.34.self_attn.o_proj.scales": "model.safetensors",
|
| 730 |
+
"model.layers.34.self_attn.o_proj.weight": "model.safetensors",
|
| 731 |
+
"model.layers.34.self_attn.q_norm.weight": "model.safetensors",
|
| 732 |
+
"model.layers.34.self_attn.q_proj.biases": "model.safetensors",
|
| 733 |
+
"model.layers.34.self_attn.q_proj.scales": "model.safetensors",
|
| 734 |
+
"model.layers.34.self_attn.q_proj.weight": "model.safetensors",
|
| 735 |
+
"model.layers.34.self_attn.v_proj.biases": "model.safetensors",
|
| 736 |
+
"model.layers.34.self_attn.v_proj.scales": "model.safetensors",
|
| 737 |
+
"model.layers.34.self_attn.v_proj.weight": "model.safetensors",
|
| 738 |
+
"model.layers.35.input_layernorm.weight": "model.safetensors",
|
| 739 |
+
"model.layers.35.mlp.down_proj.biases": "model.safetensors",
|
| 740 |
+
"model.layers.35.mlp.down_proj.scales": "model.safetensors",
|
| 741 |
+
"model.layers.35.mlp.down_proj.weight": "model.safetensors",
|
| 742 |
+
"model.layers.35.mlp.gate_proj.biases": "model.safetensors",
|
| 743 |
+
"model.layers.35.mlp.gate_proj.scales": "model.safetensors",
|
| 744 |
+
"model.layers.35.mlp.gate_proj.weight": "model.safetensors",
|
| 745 |
+
"model.layers.35.mlp.up_proj.biases": "model.safetensors",
|
| 746 |
+
"model.layers.35.mlp.up_proj.scales": "model.safetensors",
|
| 747 |
+
"model.layers.35.mlp.up_proj.weight": "model.safetensors",
|
| 748 |
+
"model.layers.35.post_attention_layernorm.weight": "model.safetensors",
|
| 749 |
+
"model.layers.35.self_attn.k_norm.weight": "model.safetensors",
|
| 750 |
+
"model.layers.35.self_attn.k_proj.biases": "model.safetensors",
|
| 751 |
+
"model.layers.35.self_attn.k_proj.scales": "model.safetensors",
|
| 752 |
+
"model.layers.35.self_attn.k_proj.weight": "model.safetensors",
|
| 753 |
+
"model.layers.35.self_attn.o_proj.biases": "model.safetensors",
|
| 754 |
+
"model.layers.35.self_attn.o_proj.scales": "model.safetensors",
|
| 755 |
+
"model.layers.35.self_attn.o_proj.weight": "model.safetensors",
|
| 756 |
+
"model.layers.35.self_attn.q_norm.weight": "model.safetensors",
|
| 757 |
+
"model.layers.35.self_attn.q_proj.biases": "model.safetensors",
|
| 758 |
+
"model.layers.35.self_attn.q_proj.scales": "model.safetensors",
|
| 759 |
+
"model.layers.35.self_attn.q_proj.weight": "model.safetensors",
|
| 760 |
+
"model.layers.35.self_attn.v_proj.biases": "model.safetensors",
|
| 761 |
+
"model.layers.35.self_attn.v_proj.scales": "model.safetensors",
|
| 762 |
+
"model.layers.35.self_attn.v_proj.weight": "model.safetensors",
|
| 763 |
+
"model.layers.4.input_layernorm.weight": "model.safetensors",
|
| 764 |
+
"model.layers.4.mlp.down_proj.biases": "model.safetensors",
|
| 765 |
+
"model.layers.4.mlp.down_proj.scales": "model.safetensors",
|
| 766 |
+
"model.layers.4.mlp.down_proj.weight": "model.safetensors",
|
| 767 |
+
"model.layers.4.mlp.gate_proj.biases": "model.safetensors",
|
| 768 |
+
"model.layers.4.mlp.gate_proj.scales": "model.safetensors",
|
| 769 |
+
"model.layers.4.mlp.gate_proj.weight": "model.safetensors",
|
| 770 |
+
"model.layers.4.mlp.up_proj.biases": "model.safetensors",
|
| 771 |
+
"model.layers.4.mlp.up_proj.scales": "model.safetensors",
|
| 772 |
+
"model.layers.4.mlp.up_proj.weight": "model.safetensors",
|
| 773 |
+
"model.layers.4.post_attention_layernorm.weight": "model.safetensors",
|
| 774 |
+
"model.layers.4.self_attn.k_norm.weight": "model.safetensors",
|
| 775 |
+
"model.layers.4.self_attn.k_proj.biases": "model.safetensors",
|
| 776 |
+
"model.layers.4.self_attn.k_proj.scales": "model.safetensors",
|
| 777 |
+
"model.layers.4.self_attn.k_proj.weight": "model.safetensors",
|
| 778 |
+
"model.layers.4.self_attn.o_proj.biases": "model.safetensors",
|
| 779 |
+
"model.layers.4.self_attn.o_proj.scales": "model.safetensors",
|
| 780 |
+
"model.layers.4.self_attn.o_proj.weight": "model.safetensors",
|
| 781 |
+
"model.layers.4.self_attn.q_norm.weight": "model.safetensors",
|
| 782 |
+
"model.layers.4.self_attn.q_proj.biases": "model.safetensors",
|
| 783 |
+
"model.layers.4.self_attn.q_proj.scales": "model.safetensors",
|
| 784 |
+
"model.layers.4.self_attn.q_proj.weight": "model.safetensors",
|
| 785 |
+
"model.layers.4.self_attn.v_proj.biases": "model.safetensors",
|
| 786 |
+
"model.layers.4.self_attn.v_proj.scales": "model.safetensors",
|
| 787 |
+
"model.layers.4.self_attn.v_proj.weight": "model.safetensors",
|
| 788 |
+
"model.layers.5.input_layernorm.weight": "model.safetensors",
|
| 789 |
+
"model.layers.5.mlp.down_proj.biases": "model.safetensors",
|
| 790 |
+
"model.layers.5.mlp.down_proj.scales": "model.safetensors",
|
| 791 |
+
"model.layers.5.mlp.down_proj.weight": "model.safetensors",
|
| 792 |
+
"model.layers.5.mlp.gate_proj.biases": "model.safetensors",
|
| 793 |
+
"model.layers.5.mlp.gate_proj.scales": "model.safetensors",
|
| 794 |
+
"model.layers.5.mlp.gate_proj.weight": "model.safetensors",
|
| 795 |
+
"model.layers.5.mlp.up_proj.biases": "model.safetensors",
|
| 796 |
+
"model.layers.5.mlp.up_proj.scales": "model.safetensors",
|
| 797 |
+
"model.layers.5.mlp.up_proj.weight": "model.safetensors",
|
| 798 |
+
"model.layers.5.post_attention_layernorm.weight": "model.safetensors",
|
| 799 |
+
"model.layers.5.self_attn.k_norm.weight": "model.safetensors",
|
| 800 |
+
"model.layers.5.self_attn.k_proj.biases": "model.safetensors",
|
| 801 |
+
"model.layers.5.self_attn.k_proj.scales": "model.safetensors",
|
| 802 |
+
"model.layers.5.self_attn.k_proj.weight": "model.safetensors",
|
| 803 |
+
"model.layers.5.self_attn.o_proj.biases": "model.safetensors",
|
| 804 |
+
"model.layers.5.self_attn.o_proj.scales": "model.safetensors",
|
| 805 |
+
"model.layers.5.self_attn.o_proj.weight": "model.safetensors",
|
| 806 |
+
"model.layers.5.self_attn.q_norm.weight": "model.safetensors",
|
| 807 |
+
"model.layers.5.self_attn.q_proj.biases": "model.safetensors",
|
| 808 |
+
"model.layers.5.self_attn.q_proj.scales": "model.safetensors",
|
| 809 |
+
"model.layers.5.self_attn.q_proj.weight": "model.safetensors",
|
| 810 |
+
"model.layers.5.self_attn.v_proj.biases": "model.safetensors",
|
| 811 |
+
"model.layers.5.self_attn.v_proj.scales": "model.safetensors",
|
| 812 |
+
"model.layers.5.self_attn.v_proj.weight": "model.safetensors",
|
| 813 |
+
"model.layers.6.input_layernorm.weight": "model.safetensors",
|
| 814 |
+
"model.layers.6.mlp.down_proj.biases": "model.safetensors",
|
| 815 |
+
"model.layers.6.mlp.down_proj.scales": "model.safetensors",
|
| 816 |
+
"model.layers.6.mlp.down_proj.weight": "model.safetensors",
|
| 817 |
+
"model.layers.6.mlp.gate_proj.biases": "model.safetensors",
|
| 818 |
+
"model.layers.6.mlp.gate_proj.scales": "model.safetensors",
|
| 819 |
+
"model.layers.6.mlp.gate_proj.weight": "model.safetensors",
|
| 820 |
+
"model.layers.6.mlp.up_proj.biases": "model.safetensors",
|
| 821 |
+
"model.layers.6.mlp.up_proj.scales": "model.safetensors",
|
| 822 |
+
"model.layers.6.mlp.up_proj.weight": "model.safetensors",
|
| 823 |
+
"model.layers.6.post_attention_layernorm.weight": "model.safetensors",
|
| 824 |
+
"model.layers.6.self_attn.k_norm.weight": "model.safetensors",
|
| 825 |
+
"model.layers.6.self_attn.k_proj.biases": "model.safetensors",
|
| 826 |
+
"model.layers.6.self_attn.k_proj.scales": "model.safetensors",
|
| 827 |
+
"model.layers.6.self_attn.k_proj.weight": "model.safetensors",
|
| 828 |
+
"model.layers.6.self_attn.o_proj.biases": "model.safetensors",
|
| 829 |
+
"model.layers.6.self_attn.o_proj.scales": "model.safetensors",
|
| 830 |
+
"model.layers.6.self_attn.o_proj.weight": "model.safetensors",
|
| 831 |
+
"model.layers.6.self_attn.q_norm.weight": "model.safetensors",
|
| 832 |
+
"model.layers.6.self_attn.q_proj.biases": "model.safetensors",
|
| 833 |
+
"model.layers.6.self_attn.q_proj.scales": "model.safetensors",
|
| 834 |
+
"model.layers.6.self_attn.q_proj.weight": "model.safetensors",
|
| 835 |
+
"model.layers.6.self_attn.v_proj.biases": "model.safetensors",
|
| 836 |
+
"model.layers.6.self_attn.v_proj.scales": "model.safetensors",
|
| 837 |
+
"model.layers.6.self_attn.v_proj.weight": "model.safetensors",
|
| 838 |
+
"model.layers.7.input_layernorm.weight": "model.safetensors",
|
| 839 |
+
"model.layers.7.mlp.down_proj.biases": "model.safetensors",
|
| 840 |
+
"model.layers.7.mlp.down_proj.scales": "model.safetensors",
|
| 841 |
+
"model.layers.7.mlp.down_proj.weight": "model.safetensors",
|
| 842 |
+
"model.layers.7.mlp.gate_proj.biases": "model.safetensors",
|
| 843 |
+
"model.layers.7.mlp.gate_proj.scales": "model.safetensors",
|
| 844 |
+
"model.layers.7.mlp.gate_proj.weight": "model.safetensors",
|
| 845 |
+
"model.layers.7.mlp.up_proj.biases": "model.safetensors",
|
| 846 |
+
"model.layers.7.mlp.up_proj.scales": "model.safetensors",
|
| 847 |
+
"model.layers.7.mlp.up_proj.weight": "model.safetensors",
|
| 848 |
+
"model.layers.7.post_attention_layernorm.weight": "model.safetensors",
|
| 849 |
+
"model.layers.7.self_attn.k_norm.weight": "model.safetensors",
|
| 850 |
+
"model.layers.7.self_attn.k_proj.biases": "model.safetensors",
|
| 851 |
+
"model.layers.7.self_attn.k_proj.scales": "model.safetensors",
|
| 852 |
+
"model.layers.7.self_attn.k_proj.weight": "model.safetensors",
|
| 853 |
+
"model.layers.7.self_attn.o_proj.biases": "model.safetensors",
|
| 854 |
+
"model.layers.7.self_attn.o_proj.scales": "model.safetensors",
|
| 855 |
+
"model.layers.7.self_attn.o_proj.weight": "model.safetensors",
|
| 856 |
+
"model.layers.7.self_attn.q_norm.weight": "model.safetensors",
|
| 857 |
+
"model.layers.7.self_attn.q_proj.biases": "model.safetensors",
|
| 858 |
+
"model.layers.7.self_attn.q_proj.scales": "model.safetensors",
|
| 859 |
+
"model.layers.7.self_attn.q_proj.weight": "model.safetensors",
|
| 860 |
+
"model.layers.7.self_attn.v_proj.biases": "model.safetensors",
|
| 861 |
+
"model.layers.7.self_attn.v_proj.scales": "model.safetensors",
|
| 862 |
+
"model.layers.7.self_attn.v_proj.weight": "model.safetensors",
|
| 863 |
+
"model.layers.8.input_layernorm.weight": "model.safetensors",
|
| 864 |
+
"model.layers.8.mlp.down_proj.biases": "model.safetensors",
|
| 865 |
+
"model.layers.8.mlp.down_proj.scales": "model.safetensors",
|
| 866 |
+
"model.layers.8.mlp.down_proj.weight": "model.safetensors",
|
| 867 |
+
"model.layers.8.mlp.gate_proj.biases": "model.safetensors",
|
| 868 |
+
"model.layers.8.mlp.gate_proj.scales": "model.safetensors",
|
| 869 |
+
"model.layers.8.mlp.gate_proj.weight": "model.safetensors",
|
| 870 |
+
"model.layers.8.mlp.up_proj.biases": "model.safetensors",
|
| 871 |
+
"model.layers.8.mlp.up_proj.scales": "model.safetensors",
|
| 872 |
+
"model.layers.8.mlp.up_proj.weight": "model.safetensors",
|
| 873 |
+
"model.layers.8.post_attention_layernorm.weight": "model.safetensors",
|
| 874 |
+
"model.layers.8.self_attn.k_norm.weight": "model.safetensors",
|
| 875 |
+
"model.layers.8.self_attn.k_proj.biases": "model.safetensors",
|
| 876 |
+
"model.layers.8.self_attn.k_proj.scales": "model.safetensors",
|
| 877 |
+
"model.layers.8.self_attn.k_proj.weight": "model.safetensors",
|
| 878 |
+
"model.layers.8.self_attn.o_proj.biases": "model.safetensors",
|
| 879 |
+
"model.layers.8.self_attn.o_proj.scales": "model.safetensors",
|
| 880 |
+
"model.layers.8.self_attn.o_proj.weight": "model.safetensors",
|
| 881 |
+
"model.layers.8.self_attn.q_norm.weight": "model.safetensors",
|
| 882 |
+
"model.layers.8.self_attn.q_proj.biases": "model.safetensors",
|
| 883 |
+
"model.layers.8.self_attn.q_proj.scales": "model.safetensors",
|
| 884 |
+
"model.layers.8.self_attn.q_proj.weight": "model.safetensors",
|
| 885 |
+
"model.layers.8.self_attn.v_proj.biases": "model.safetensors",
|
| 886 |
+
"model.layers.8.self_attn.v_proj.scales": "model.safetensors",
|
| 887 |
+
"model.layers.8.self_attn.v_proj.weight": "model.safetensors",
|
| 888 |
+
"model.layers.9.input_layernorm.weight": "model.safetensors",
|
| 889 |
+
"model.layers.9.mlp.down_proj.biases": "model.safetensors",
|
| 890 |
+
"model.layers.9.mlp.down_proj.scales": "model.safetensors",
|
| 891 |
+
"model.layers.9.mlp.down_proj.weight": "model.safetensors",
|
| 892 |
+
"model.layers.9.mlp.gate_proj.biases": "model.safetensors",
|
| 893 |
+
"model.layers.9.mlp.gate_proj.scales": "model.safetensors",
|
| 894 |
+
"model.layers.9.mlp.gate_proj.weight": "model.safetensors",
|
| 895 |
+
"model.layers.9.mlp.up_proj.biases": "model.safetensors",
|
| 896 |
+
"model.layers.9.mlp.up_proj.scales": "model.safetensors",
|
| 897 |
+
"model.layers.9.mlp.up_proj.weight": "model.safetensors",
|
| 898 |
+
"model.layers.9.post_attention_layernorm.weight": "model.safetensors",
|
| 899 |
+
"model.layers.9.self_attn.k_norm.weight": "model.safetensors",
|
| 900 |
+
"model.layers.9.self_attn.k_proj.biases": "model.safetensors",
|
| 901 |
+
"model.layers.9.self_attn.k_proj.scales": "model.safetensors",
|
| 902 |
+
"model.layers.9.self_attn.k_proj.weight": "model.safetensors",
|
| 903 |
+
"model.layers.9.self_attn.o_proj.biases": "model.safetensors",
|
| 904 |
+
"model.layers.9.self_attn.o_proj.scales": "model.safetensors",
|
| 905 |
+
"model.layers.9.self_attn.o_proj.weight": "model.safetensors",
|
| 906 |
+
"model.layers.9.self_attn.q_norm.weight": "model.safetensors",
|
| 907 |
+
"model.layers.9.self_attn.q_proj.biases": "model.safetensors",
|
| 908 |
+
"model.layers.9.self_attn.q_proj.scales": "model.safetensors",
|
| 909 |
+
"model.layers.9.self_attn.q_proj.weight": "model.safetensors",
|
| 910 |
+
"model.layers.9.self_attn.v_proj.biases": "model.safetensors",
|
| 911 |
+
"model.layers.9.self_attn.v_proj.scales": "model.safetensors",
|
| 912 |
+
"model.layers.9.self_attn.v_proj.weight": "model.safetensors",
|
| 913 |
+
"model.norm.weight": "model.safetensors"
|
| 914 |
+
}
|
| 915 |
+
}
|
mlx_llm/special_tokens_map.json
ADDED
|
@@ -0,0 +1,31 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"additional_special_tokens": [
|
| 3 |
+
"<|im_start|>",
|
| 4 |
+
"<|im_end|>",
|
| 5 |
+
"<|object_ref_start|>",
|
| 6 |
+
"<|object_ref_end|>",
|
| 7 |
+
"<|box_start|>",
|
| 8 |
+
"<|box_end|>",
|
| 9 |
+
"<|quad_start|>",
|
| 10 |
+
"<|quad_end|>",
|
| 11 |
+
"<|vision_start|>",
|
| 12 |
+
"<|vision_end|>",
|
| 13 |
+
"<|vision_pad|>",
|
| 14 |
+
"<|image_pad|>",
|
| 15 |
+
"<|video_pad|>"
|
| 16 |
+
],
|
| 17 |
+
"eos_token": {
|
| 18 |
+
"content": "<|im_end|>",
|
| 19 |
+
"lstrip": false,
|
| 20 |
+
"normalized": false,
|
| 21 |
+
"rstrip": false,
|
| 22 |
+
"single_word": false
|
| 23 |
+
},
|
| 24 |
+
"pad_token": {
|
| 25 |
+
"content": "<|endoftext|>",
|
| 26 |
+
"lstrip": false,
|
| 27 |
+
"normalized": false,
|
| 28 |
+
"rstrip": false,
|
| 29 |
+
"single_word": false
|
| 30 |
+
}
|
| 31 |
+
}
|
mlx_llm/tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9e4855778eb5536cbc88a3ad243c339bce177b610474b2305dc1bdf322c65482
|
| 3 |
+
size 11423398
|
mlx_llm/tokenizer_config.json
ADDED
|
@@ -0,0 +1,271 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_bos_token": false,
|
| 3 |
+
"add_prefix_space": false,
|
| 4 |
+
"added_tokens_decoder": {
|
| 5 |
+
"151643": {
|
| 6 |
+
"content": "<|endoftext|>",
|
| 7 |
+
"lstrip": false,
|
| 8 |
+
"normalized": false,
|
| 9 |
+
"rstrip": false,
|
| 10 |
+
"single_word": false,
|
| 11 |
+
"special": true
|
| 12 |
+
},
|
| 13 |
+
"151644": {
|
| 14 |
+
"content": "<|im_start|>",
|
| 15 |
+
"lstrip": false,
|
| 16 |
+
"normalized": false,
|
| 17 |
+
"rstrip": false,
|
| 18 |
+
"single_word": false,
|
| 19 |
+
"special": true
|
| 20 |
+
},
|
| 21 |
+
"151645": {
|
| 22 |
+
"content": "<|im_end|>",
|
| 23 |
+
"lstrip": false,
|
| 24 |
+
"normalized": false,
|
| 25 |
+
"rstrip": false,
|
| 26 |
+
"single_word": false,
|
| 27 |
+
"special": true
|
| 28 |
+
},
|
| 29 |
+
"151646": {
|
| 30 |
+
"content": "<|object_ref_start|>",
|
| 31 |
+
"lstrip": false,
|
| 32 |
+
"normalized": false,
|
| 33 |
+
"rstrip": false,
|
| 34 |
+
"single_word": false,
|
| 35 |
+
"special": true
|
| 36 |
+
},
|
| 37 |
+
"151647": {
|
| 38 |
+
"content": "<|object_ref_end|>",
|
| 39 |
+
"lstrip": false,
|
| 40 |
+
"normalized": false,
|
| 41 |
+
"rstrip": false,
|
| 42 |
+
"single_word": false,
|
| 43 |
+
"special": true
|
| 44 |
+
},
|
| 45 |
+
"151648": {
|
| 46 |
+
"content": "<|box_start|>",
|
| 47 |
+
"lstrip": false,
|
| 48 |
+
"normalized": false,
|
| 49 |
+
"rstrip": false,
|
| 50 |
+
"single_word": false,
|
| 51 |
+
"special": true
|
| 52 |
+
},
|
| 53 |
+
"151649": {
|
| 54 |
+
"content": "<|box_end|>",
|
| 55 |
+
"lstrip": false,
|
| 56 |
+
"normalized": false,
|
| 57 |
+
"rstrip": false,
|
| 58 |
+
"single_word": false,
|
| 59 |
+
"special": true
|
| 60 |
+
},
|
| 61 |
+
"151650": {
|
| 62 |
+
"content": "<|quad_start|>",
|
| 63 |
+
"lstrip": false,
|
| 64 |
+
"normalized": false,
|
| 65 |
+
"rstrip": false,
|
| 66 |
+
"single_word": false,
|
| 67 |
+
"special": true
|
| 68 |
+
},
|
| 69 |
+
"151651": {
|
| 70 |
+
"content": "<|quad_end|>",
|
| 71 |
+
"lstrip": false,
|
| 72 |
+
"normalized": false,
|
| 73 |
+
"rstrip": false,
|
| 74 |
+
"single_word": false,
|
| 75 |
+
"special": true
|
| 76 |
+
},
|
| 77 |
+
"151652": {
|
| 78 |
+
"content": "<|vision_start|>",
|
| 79 |
+
"lstrip": false,
|
| 80 |
+
"normalized": false,
|
| 81 |
+
"rstrip": false,
|
| 82 |
+
"single_word": false,
|
| 83 |
+
"special": true
|
| 84 |
+
},
|
| 85 |
+
"151653": {
|
| 86 |
+
"content": "<|vision_end|>",
|
| 87 |
+
"lstrip": false,
|
| 88 |
+
"normalized": false,
|
| 89 |
+
"rstrip": false,
|
| 90 |
+
"single_word": false,
|
| 91 |
+
"special": true
|
| 92 |
+
},
|
| 93 |
+
"151654": {
|
| 94 |
+
"content": "<|vision_pad|>",
|
| 95 |
+
"lstrip": false,
|
| 96 |
+
"normalized": false,
|
| 97 |
+
"rstrip": false,
|
| 98 |
+
"single_word": false,
|
| 99 |
+
"special": true
|
| 100 |
+
},
|
| 101 |
+
"151655": {
|
| 102 |
+
"content": "<|image_pad|>",
|
| 103 |
+
"lstrip": false,
|
| 104 |
+
"normalized": false,
|
| 105 |
+
"rstrip": false,
|
| 106 |
+
"single_word": false,
|
| 107 |
+
"special": true
|
| 108 |
+
},
|
| 109 |
+
"151656": {
|
| 110 |
+
"content": "<|video_pad|>",
|
| 111 |
+
"lstrip": false,
|
| 112 |
+
"normalized": false,
|
| 113 |
+
"rstrip": false,
|
| 114 |
+
"single_word": false,
|
| 115 |
+
"special": true
|
| 116 |
+
},
|
| 117 |
+
"151657": {
|
| 118 |
+
"content": "<tool_call>",
|
| 119 |
+
"lstrip": false,
|
| 120 |
+
"normalized": false,
|
| 121 |
+
"rstrip": false,
|
| 122 |
+
"single_word": false,
|
| 123 |
+
"special": false
|
| 124 |
+
},
|
| 125 |
+
"151658": {
|
| 126 |
+
"content": "</tool_call>",
|
| 127 |
+
"lstrip": false,
|
| 128 |
+
"normalized": false,
|
| 129 |
+
"rstrip": false,
|
| 130 |
+
"single_word": false,
|
| 131 |
+
"special": false
|
| 132 |
+
},
|
| 133 |
+
"151659": {
|
| 134 |
+
"content": "<|fim_prefix|>",
|
| 135 |
+
"lstrip": false,
|
| 136 |
+
"normalized": false,
|
| 137 |
+
"rstrip": false,
|
| 138 |
+
"single_word": false,
|
| 139 |
+
"special": false
|
| 140 |
+
},
|
| 141 |
+
"151660": {
|
| 142 |
+
"content": "<|fim_middle|>",
|
| 143 |
+
"lstrip": false,
|
| 144 |
+
"normalized": false,
|
| 145 |
+
"rstrip": false,
|
| 146 |
+
"single_word": false,
|
| 147 |
+
"special": false
|
| 148 |
+
},
|
| 149 |
+
"151661": {
|
| 150 |
+
"content": "<|fim_suffix|>",
|
| 151 |
+
"lstrip": false,
|
| 152 |
+
"normalized": false,
|
| 153 |
+
"rstrip": false,
|
| 154 |
+
"single_word": false,
|
| 155 |
+
"special": false
|
| 156 |
+
},
|
| 157 |
+
"151662": {
|
| 158 |
+
"content": "<|fim_pad|>",
|
| 159 |
+
"lstrip": false,
|
| 160 |
+
"normalized": false,
|
| 161 |
+
"rstrip": false,
|
| 162 |
+
"single_word": false,
|
| 163 |
+
"special": false
|
| 164 |
+
},
|
| 165 |
+
"151663": {
|
| 166 |
+
"content": "<|repo_name|>",
|
| 167 |
+
"lstrip": false,
|
| 168 |
+
"normalized": false,
|
| 169 |
+
"rstrip": false,
|
| 170 |
+
"single_word": false,
|
| 171 |
+
"special": false
|
| 172 |
+
},
|
| 173 |
+
"151664": {
|
| 174 |
+
"content": "<|file_sep|>",
|
| 175 |
+
"lstrip": false,
|
| 176 |
+
"normalized": false,
|
| 177 |
+
"rstrip": false,
|
| 178 |
+
"single_word": false,
|
| 179 |
+
"special": false
|
| 180 |
+
},
|
| 181 |
+
"151665": {
|
| 182 |
+
"content": "<tool_response>",
|
| 183 |
+
"lstrip": false,
|
| 184 |
+
"normalized": false,
|
| 185 |
+
"rstrip": false,
|
| 186 |
+
"single_word": false,
|
| 187 |
+
"special": false
|
| 188 |
+
},
|
| 189 |
+
"151666": {
|
| 190 |
+
"content": "</tool_response>",
|
| 191 |
+
"lstrip": false,
|
| 192 |
+
"normalized": false,
|
| 193 |
+
"rstrip": false,
|
| 194 |
+
"single_word": false,
|
| 195 |
+
"special": false
|
| 196 |
+
},
|
| 197 |
+
"151667": {
|
| 198 |
+
"content": "<think>",
|
| 199 |
+
"lstrip": false,
|
| 200 |
+
"normalized": false,
|
| 201 |
+
"rstrip": false,
|
| 202 |
+
"single_word": false,
|
| 203 |
+
"special": false
|
| 204 |
+
},
|
| 205 |
+
"151668": {
|
| 206 |
+
"content": "</think>",
|
| 207 |
+
"lstrip": false,
|
| 208 |
+
"normalized": false,
|
| 209 |
+
"rstrip": false,
|
| 210 |
+
"single_word": false,
|
| 211 |
+
"special": false
|
| 212 |
+
},
|
| 213 |
+
"151669": {
|
| 214 |
+
"content": "<|system|>",
|
| 215 |
+
"lstrip": false,
|
| 216 |
+
"normalized": false,
|
| 217 |
+
"rstrip": false,
|
| 218 |
+
"single_word": false,
|
| 219 |
+
"special": false
|
| 220 |
+
},
|
| 221 |
+
"151670": {
|
| 222 |
+
"content": "<|user|>",
|
| 223 |
+
"lstrip": false,
|
| 224 |
+
"normalized": false,
|
| 225 |
+
"rstrip": false,
|
| 226 |
+
"single_word": false,
|
| 227 |
+
"special": false
|
| 228 |
+
},
|
| 229 |
+
"151671": {
|
| 230 |
+
"content": "<|assistant|>",
|
| 231 |
+
"lstrip": false,
|
| 232 |
+
"normalized": false,
|
| 233 |
+
"rstrip": false,
|
| 234 |
+
"single_word": false,
|
| 235 |
+
"special": false
|
| 236 |
+
},
|
| 237 |
+
"151672": {
|
| 238 |
+
"content": "<|eot|>",
|
| 239 |
+
"lstrip": false,
|
| 240 |
+
"normalized": false,
|
| 241 |
+
"rstrip": false,
|
| 242 |
+
"single_word": false,
|
| 243 |
+
"special": false
|
| 244 |
+
}
|
| 245 |
+
},
|
| 246 |
+
"additional_special_tokens": [
|
| 247 |
+
"<|im_start|>",
|
| 248 |
+
"<|im_end|>",
|
| 249 |
+
"<|object_ref_start|>",
|
| 250 |
+
"<|object_ref_end|>",
|
| 251 |
+
"<|box_start|>",
|
| 252 |
+
"<|box_end|>",
|
| 253 |
+
"<|quad_start|>",
|
| 254 |
+
"<|quad_end|>",
|
| 255 |
+
"<|vision_start|>",
|
| 256 |
+
"<|vision_end|>",
|
| 257 |
+
"<|vision_pad|>",
|
| 258 |
+
"<|image_pad|>",
|
| 259 |
+
"<|video_pad|>"
|
| 260 |
+
],
|
| 261 |
+
"bos_token": null,
|
| 262 |
+
"clean_up_tokenization_spaces": false,
|
| 263 |
+
"eos_token": "<|im_end|>",
|
| 264 |
+
"errors": "replace",
|
| 265 |
+
"extra_special_tokens": {},
|
| 266 |
+
"model_max_length": 131072,
|
| 267 |
+
"pad_token": "<|endoftext|>",
|
| 268 |
+
"split_special_tokens": false,
|
| 269 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 270 |
+
"unk_token": null
|
| 271 |
+
}
|
mlx_llm/vocab.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
run_moss.py
ADDED
|
@@ -0,0 +1,94 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""MOSS-Audio MLX runner for /srt eval — soundfile (no numba) + ASR transcription prompt + time markers.
|
| 2 |
+
Usage: python run_moss.py --model <dir> --audio clip.wav [--prompt "..."] [--max-tokens 4096] [--temp 0.0]
|
| 3 |
+
Works for both 4B-Thinking and 8B-hybrid bundles (audio int4 auto-detected).
|
| 4 |
+
"""
|
| 5 |
+
from __future__ import annotations
|
| 6 |
+
import argparse, sys, time
|
| 7 |
+
from pathlib import Path
|
| 8 |
+
|
| 9 |
+
def main():
|
| 10 |
+
p = argparse.ArgumentParser()
|
| 11 |
+
p.add_argument("--model", required=True, help="bundle dir (has mlx_llm/ mlx_audio/ scripts/)")
|
| 12 |
+
p.add_argument("--audio", required=True)
|
| 13 |
+
p.add_argument("--prompt", default="Please transcribe this audio.")
|
| 14 |
+
p.add_argument("--max-tokens", type=int, default=4096)
|
| 15 |
+
p.add_argument("--temp", type=float, default=0.0)
|
| 16 |
+
p.add_argument("--repetition-penalty", type=float, default=1.02)
|
| 17 |
+
p.add_argument("--no-time-marker", action="store_true")
|
| 18 |
+
args = p.parse_args()
|
| 19 |
+
|
| 20 |
+
HERE = Path(args.model).resolve()
|
| 21 |
+
sys.path.insert(0, str(HERE / "scripts"))
|
| 22 |
+
import soundfile as sf
|
| 23 |
+
import mlx.core as mx
|
| 24 |
+
import numpy as np
|
| 25 |
+
from mlx_lm import load as mlx_load
|
| 26 |
+
from mlx_lm.generate import generate_step
|
| 27 |
+
from mlx_lm.sample_utils import make_sampler, make_logits_processors
|
| 28 |
+
from moss_audio_mlx_bridge_v3 import load_mlx_audio_path, run_mlx_audio_pipeline, install_deepstack_hooks
|
| 29 |
+
from moss_audio_mel_mlx import build_mel_and_input_ids
|
| 30 |
+
|
| 31 |
+
ad_w = mx.load(str(HERE / "mlx_audio/audio_adapter.safetensors"))
|
| 32 |
+
int4_audio = "down_proj.scales" in ad_w
|
| 33 |
+
llm_hidden = (ad_w["down_proj.scales"].shape[0] if int4_audio else ad_w["down_proj.weight"].shape[0])
|
| 34 |
+
size_tag = "4B" if llm_hidden == 2560 else "8B"
|
| 35 |
+
print(f"[detect] {size_tag} bundle, audio_int4={int4_audio}, prompt={args.prompt!r}, time_marker={not args.no_time_marker}", flush=True)
|
| 36 |
+
|
| 37 |
+
t0 = time.perf_counter()
|
| 38 |
+
mlx_model, tok = mlx_load(str(HERE / "mlx_llm"))
|
| 39 |
+
print(f"[load] LLM {time.perf_counter()-t0:.1f}s peak={mx.get_peak_memory()/1e9:.2f}GB", flush=True)
|
| 40 |
+
t0 = time.perf_counter()
|
| 41 |
+
encoder, adapter, mergers = load_mlx_audio_path(HERE / "mlx_audio", int4=int4_audio)
|
| 42 |
+
print(f"[load] audio path {time.perf_counter()-t0:.1f}s", flush=True)
|
| 43 |
+
|
| 44 |
+
y, sr = sf.read(args.audio)
|
| 45 |
+
y = np.asarray(y, dtype=np.float32)
|
| 46 |
+
if y.ndim > 1:
|
| 47 |
+
y = y.mean(axis=1)
|
| 48 |
+
assert sr == 16000, f"expected 16kHz, got {sr}"
|
| 49 |
+
print(f"[audio] {args.audio} ({len(y)/16000:.1f}s)", flush=True)
|
| 50 |
+
|
| 51 |
+
t0 = time.perf_counter()
|
| 52 |
+
mel, lens, input_ids_mx, audio_token_id = build_mel_and_input_ids(
|
| 53 |
+
y, tok, prompt=args.prompt, enable_time_marker=not args.no_time_marker)
|
| 54 |
+
primary, ds_embeds = run_mlx_audio_pipeline(encoder, adapter, mergers, mel, lens)
|
| 55 |
+
primary = primary.astype(mx.bfloat16)
|
| 56 |
+
ds_embeds = [d.astype(mx.bfloat16) for d in ds_embeds]
|
| 57 |
+
mx.eval(primary, *ds_embeds)
|
| 58 |
+
print(f"[encode] {time.perf_counter()-t0:.2f}s primary={primary.shape}", flush=True)
|
| 59 |
+
del encoder, adapter, mergers, mel, lens
|
| 60 |
+
import gc; gc.collect(); mx.clear_cache()
|
| 61 |
+
|
| 62 |
+
audio_mask = input_ids_mx == audio_token_id
|
| 63 |
+
audio_positions = np.where(np.array(audio_mask[0]))[0]
|
| 64 |
+
text_embeds = mlx_model.model.embed_tokens(input_ids_mx)
|
| 65 |
+
text_np = np.array(text_embeds.astype(mx.float32))
|
| 66 |
+
primary_np = np.array(primary.astype(mx.float32))
|
| 67 |
+
text_np[0, audio_positions, :] = primary_np[0, :, :]
|
| 68 |
+
merged = mx.array(text_np).astype(mx.bfloat16)
|
| 69 |
+
ds_flat = [d[0] for d in ds_embeds]
|
| 70 |
+
install_deepstack_hooks(mlx_model, ds_flat, audio_positions)
|
| 71 |
+
|
| 72 |
+
sampler = make_sampler(temp=args.temp, top_p=1.0, top_k=(0 if args.temp == 0 else 50))
|
| 73 |
+
gen_kwargs = dict(prompt=input_ids_mx[0], model=mlx_model,
|
| 74 |
+
input_embeddings=merged[0], max_tokens=args.max_tokens, sampler=sampler)
|
| 75 |
+
if args.repetition_penalty:
|
| 76 |
+
gen_kwargs["logits_processors"] = make_logits_processors(
|
| 77 |
+
repetition_penalty=args.repetition_penalty, repetition_context_size=20)
|
| 78 |
+
|
| 79 |
+
t0 = time.perf_counter(); generated = []; ttft = None
|
| 80 |
+
for tok_id, _ in generate_step(**gen_kwargs):
|
| 81 |
+
if ttft is None:
|
| 82 |
+
ttft = time.perf_counter() - t0
|
| 83 |
+
generated.append(int(tok_id))
|
| 84 |
+
if tok_id == tok.eos_token_id:
|
| 85 |
+
break
|
| 86 |
+
elapsed = time.perf_counter() - t0
|
| 87 |
+
n = len(generated)
|
| 88 |
+
decode_s = max(elapsed - (ttft or 0), 1e-6)
|
| 89 |
+
print(f"[gen] {n}tok in {elapsed:.1f}s ttft={ttft:.2f}s decode={max(n-1,0)/decode_s:.1f}t/s "
|
| 90 |
+
f"rtf={(len(y)/16000)/elapsed:.2f}x peak={mx.get_peak_memory()/1e9:.2f}GB", flush=True)
|
| 91 |
+
print(f"\n=== OUTPUT ===\n{tok.decode(generated)}\n=== END ===", flush=True)
|
| 92 |
+
|
| 93 |
+
if __name__ == "__main__":
|
| 94 |
+
main()
|
scripts/__pycache__/moss_audio_encoder_mlx.cpython-312.pyc
ADDED
|
Binary file (13.5 kB). View file
|
|
|
scripts/__pycache__/moss_audio_mel_mlx.cpython-312.pyc
ADDED
|
Binary file (9.35 kB). View file
|
|
|
scripts/__pycache__/moss_audio_mlx_bridge_v3.cpython-312.pyc
ADDED
|
Binary file (18.8 kB). View file
|
|
|
scripts/assets/mel_filters.npz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7450ae70723a5ef9d341e3cee628c7cb0177f36ce42c44b7ed2bf3325f0f6d4c
|
| 3 |
+
size 4271
|
scripts/moss_audio_encoder_mlx.py
ADDED
|
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""MLX-native MossAudioEncoder.
|
| 2 |
+
|
| 3 |
+
Direct port of src/modeling_moss_audio.py:36-155 (MossAudioEncoder).
|
| 4 |
+
Adapted from ml-explore/mlx-examples/whisper/mlx_whisper/whisper.py with:
|
| 5 |
+
- 3× Conv2d stride-2 stem (instead of Whisper's 2× Conv1d)
|
| 6 |
+
- Pre-existing HF Whisper attribute names (q_proj/k_proj/v_proj/out_proj, fc1/fc2,
|
| 7 |
+
self_attn_layer_norm/final_layer_norm) so weight remap is near-identity
|
| 8 |
+
- DeepStack taps: capture hidden state AFTER layers in deepstack_layer_indexes
|
| 9 |
+
- feature_lens-based padding mask
|
| 10 |
+
"""
|
| 11 |
+
from __future__ import annotations
|
| 12 |
+
|
| 13 |
+
import math
|
| 14 |
+
from dataclasses import dataclass, field
|
| 15 |
+
from typing import List, Optional, Tuple
|
| 16 |
+
|
| 17 |
+
import mlx.core as mx
|
| 18 |
+
import mlx.nn as nn
|
| 19 |
+
|
| 20 |
+
|
| 21 |
+
# ---- helpers ----------------------------------------------------------
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
def sinusoids(length: int, channels: int, max_timescale: float = 10000.0) -> mx.array:
|
| 25 |
+
"""Whisper-style sinusoidal position embeddings. Matches mlx-examples whisper."""
|
| 26 |
+
assert channels % 2 == 0
|
| 27 |
+
log_timescale_increment = math.log(max_timescale) / (channels // 2 - 1)
|
| 28 |
+
inv_timescales = mx.exp(-log_timescale_increment * mx.arange(channels // 2))
|
| 29 |
+
scaled_time = mx.arange(length)[:, None] * inv_timescales[None, :]
|
| 30 |
+
return mx.concatenate([mx.sin(scaled_time), mx.cos(scaled_time)], axis=1)
|
| 31 |
+
|
| 32 |
+
|
| 33 |
+
# ---- attention ------------------------------------------------------
|
| 34 |
+
|
| 35 |
+
|
| 36 |
+
class WhisperAttention(nn.Module):
|
| 37 |
+
"""HF-Whisper-style self-attention. Layer-scaling convention (`1/sqrt(head_dim)`
|
| 38 |
+
applied to Q, not split between Q and K like mlx-examples does).
|
| 39 |
+
|
| 40 |
+
Attribute names match HF so weight remap is identity: q_proj/k_proj/v_proj/out_proj.
|
| 41 |
+
"""
|
| 42 |
+
|
| 43 |
+
def __init__(self, d_model: int, n_heads: int):
|
| 44 |
+
super().__init__()
|
| 45 |
+
self.n_heads = n_heads
|
| 46 |
+
self.head_dim = d_model // n_heads
|
| 47 |
+
assert d_model == self.head_dim * n_heads
|
| 48 |
+
# HF Whisper: q/v/out have bias; k does not
|
| 49 |
+
self.q_proj = nn.Linear(d_model, d_model, bias=True)
|
| 50 |
+
self.k_proj = nn.Linear(d_model, d_model, bias=False)
|
| 51 |
+
self.v_proj = nn.Linear(d_model, d_model, bias=True)
|
| 52 |
+
self.out_proj = nn.Linear(d_model, d_model, bias=True)
|
| 53 |
+
|
| 54 |
+
def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
|
| 55 |
+
B, T, D = x.shape
|
| 56 |
+
q = self.q_proj(x).reshape(B, T, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
|
| 57 |
+
k = self.k_proj(x).reshape(B, T, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
|
| 58 |
+
v = self.v_proj(x).reshape(B, T, self.n_heads, self.head_dim).transpose(0, 2, 1, 3)
|
| 59 |
+
scale = self.head_dim ** -0.5
|
| 60 |
+
attn = (q * scale) @ k.transpose(0, 1, 3, 2) # (B, H, T, T)
|
| 61 |
+
if mask is not None:
|
| 62 |
+
attn = attn + mask
|
| 63 |
+
w = mx.softmax(attn, axis=-1, precise=True)
|
| 64 |
+
out = (w @ v).transpose(0, 2, 1, 3).reshape(B, T, D)
|
| 65 |
+
return self.out_proj(out)
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
# ---- encoder layer --------------------------------------------------
|
| 69 |
+
|
| 70 |
+
|
| 71 |
+
class WhisperEncoderBlock(nn.Module):
|
| 72 |
+
"""Pre-LN Whisper encoder block. Matches transformers.WhisperEncoderLayer."""
|
| 73 |
+
|
| 74 |
+
def __init__(self, d_model: int, n_heads: int, ffn_dim: int):
|
| 75 |
+
super().__init__()
|
| 76 |
+
self.self_attn = WhisperAttention(d_model, n_heads)
|
| 77 |
+
self.self_attn_layer_norm = nn.LayerNorm(d_model)
|
| 78 |
+
self.fc1 = nn.Linear(d_model, ffn_dim)
|
| 79 |
+
self.fc2 = nn.Linear(ffn_dim, d_model)
|
| 80 |
+
self.final_layer_norm = nn.LayerNorm(d_model)
|
| 81 |
+
|
| 82 |
+
def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
|
| 83 |
+
h = self.self_attn_layer_norm(x)
|
| 84 |
+
x = x + self.self_attn(h, mask=mask)
|
| 85 |
+
h = self.final_layer_norm(x)
|
| 86 |
+
x = x + self.fc2(nn.gelu(self.fc1(h)))
|
| 87 |
+
return x
|
| 88 |
+
|
| 89 |
+
|
| 90 |
+
# ---- encoder --------------------------------------------------------
|
| 91 |
+
|
| 92 |
+
|
| 93 |
+
@dataclass
|
| 94 |
+
class EncoderConfig:
|
| 95 |
+
num_mel_bins: int = 128
|
| 96 |
+
downsample_hidden_size: int = 480
|
| 97 |
+
d_model: int = 1280
|
| 98 |
+
n_heads: int = 20
|
| 99 |
+
ffn_dim: int = 5120
|
| 100 |
+
n_layers: int = 32
|
| 101 |
+
max_source_positions: int = 1500
|
| 102 |
+
layer_norm_eps: float = 1e-5
|
| 103 |
+
output_dim: int = 1280
|
| 104 |
+
deepstack_layer_indexes: List[int] = field(default_factory=lambda: [8, 16, 24])
|
| 105 |
+
|
| 106 |
+
|
| 107 |
+
class MossAudioEncoderMLX(nn.Module):
|
| 108 |
+
def __init__(self, cfg: EncoderConfig):
|
| 109 |
+
super().__init__()
|
| 110 |
+
self.cfg = cfg
|
| 111 |
+
# Conv2d stem: 1 → 480 → 480 → 480, each stride-2
|
| 112 |
+
# MLX Conv2d expects NHWC, weight shape (OC, kH, kW, IC)
|
| 113 |
+
self.conv1 = nn.Conv2d(1, cfg.downsample_hidden_size, kernel_size=3, stride=2, padding=1)
|
| 114 |
+
self.conv2 = nn.Conv2d(cfg.downsample_hidden_size, cfg.downsample_hidden_size, kernel_size=3, stride=2, padding=1)
|
| 115 |
+
self.conv3 = nn.Conv2d(cfg.downsample_hidden_size, cfg.downsample_hidden_size, kernel_size=3, stride=2, padding=1)
|
| 116 |
+
# After 3× stride-2 on mel-axis (128→64→32→16): flat dim = 480*16 = 7680
|
| 117 |
+
self.stem_proj = nn.Linear(cfg.downsample_hidden_size * 16, cfg.d_model)
|
| 118 |
+
# Precomputed sinusoids, will be sliced
|
| 119 |
+
self._positions = sinusoids(cfg.max_source_positions, cfg.d_model)
|
| 120 |
+
self.layers = [
|
| 121 |
+
WhisperEncoderBlock(cfg.d_model, cfg.n_heads, cfg.ffn_dim)
|
| 122 |
+
for _ in range(cfg.n_layers)
|
| 123 |
+
]
|
| 124 |
+
self.layer_norm = nn.LayerNorm(cfg.d_model, eps=cfg.layer_norm_eps)
|
| 125 |
+
# MOSS has optional out_proj; for 4B output_dim==d_model, so it's Identity in PyTorch
|
| 126 |
+
# We skip it entirely (equivalent).
|
| 127 |
+
assert cfg.output_dim == cfg.d_model, "non-identity out_proj not yet implemented"
|
| 128 |
+
self._deepstack_set = set(cfg.deepstack_layer_indexes)
|
| 129 |
+
|
| 130 |
+
def _compute_downsampled_length(self, L: int) -> int:
|
| 131 |
+
"""3× stride-2 conv output length: ceil((((L-1)//2+1)-1)//2+1 ... )"""
|
| 132 |
+
def step(n): return (n - 1) // 2 + 1
|
| 133 |
+
return step(step(step(L)))
|
| 134 |
+
|
| 135 |
+
def __call__(
|
| 136 |
+
self,
|
| 137 |
+
input_features: mx.array, # (B, n_mels, T) bf16 mel spectrogram
|
| 138 |
+
feature_lens: Optional[mx.array] = None,
|
| 139 |
+
return_deepstack: bool = True,
|
| 140 |
+
) -> Tuple[mx.array, Optional[List[mx.array]]]:
|
| 141 |
+
if input_features.ndim == 2:
|
| 142 |
+
input_features = input_features[None]
|
| 143 |
+
B, n_mels, T = input_features.shape
|
| 144 |
+
if feature_lens is None:
|
| 145 |
+
feature_lens = mx.full((B,), T, dtype=mx.int32)
|
| 146 |
+
|
| 147 |
+
# (B, n_mels, T) → (B, n_mels, T, 1) [NHWC with channels-last = 1 input channel]
|
| 148 |
+
# But MLX Conv2d expects input shape (B, H, W, C_in). We map:
|
| 149 |
+
# H = n_mels (128), W = T (frames), C_in = 1
|
| 150 |
+
x = input_features[..., None] # (B, n_mels, T, 1)
|
| 151 |
+
x = nn.gelu(self.conv1(x)) # (B, 64, T/2, 480)
|
| 152 |
+
x = nn.gelu(self.conv2(x)) # (B, 32, T/4, 480)
|
| 153 |
+
x = nn.gelu(self.conv3(x)) # (B, 16, T/8, 480)
|
| 154 |
+
# PyTorch reference: (B, C, F, T) → permute(0,3,1,2) → (B, T, C, F) → flatten → (B, T, C*F)
|
| 155 |
+
# MLX is (B, F, T, C) post-conv. Need transpose to (B, T, C, F) to match PT's flatten order.
|
| 156 |
+
B_, H_, W_, C_ = x.shape # H_=F, W_=T, C_=C
|
| 157 |
+
x = x.transpose(0, 2, 3, 1).reshape(B_, W_, C_ * H_) # (B, T, C*F)
|
| 158 |
+
x = self.stem_proj(x) # (B, T', d_model)
|
| 159 |
+
|
| 160 |
+
# Trim to actual downsampled length (in case input was padded)
|
| 161 |
+
max_len = self._compute_downsampled_length(int(feature_lens.max().item()))
|
| 162 |
+
if x.shape[1] > max_len:
|
| 163 |
+
x = x[:, :max_len, :]
|
| 164 |
+
|
| 165 |
+
# Add sinusoidal positions
|
| 166 |
+
seq_len = x.shape[1]
|
| 167 |
+
pos = self._positions[:seq_len].astype(x.dtype)
|
| 168 |
+
x = x + pos
|
| 169 |
+
|
| 170 |
+
# Build attention mask: (B, 1, 1, seq_len) additive
|
| 171 |
+
# padding_mask[b, t] = True if t >= downsampled_len[b] (this is where we mask out)
|
| 172 |
+
dsl = mx.stack([
|
| 173 |
+
mx.array(self._compute_downsampled_length(int(feature_lens[b].item())), dtype=mx.int32)
|
| 174 |
+
for b in range(B)
|
| 175 |
+
]) # (B,)
|
| 176 |
+
ar = mx.arange(seq_len, dtype=mx.int32)
|
| 177 |
+
padding = ar[None, :] >= dsl[:, None] # (B, seq_len) bool
|
| 178 |
+
neg_inf = mx.array(-1e9, dtype=x.dtype)
|
| 179 |
+
mask = mx.where(padding, neg_inf, mx.array(0.0, dtype=x.dtype))
|
| 180 |
+
mask = mask[:, None, None, :] # (B, 1, 1, seq_len)
|
| 181 |
+
|
| 182 |
+
deepstack: List[mx.array] = []
|
| 183 |
+
for layer_idx, layer in enumerate(self.layers):
|
| 184 |
+
x = layer(x, mask=mask)
|
| 185 |
+
if return_deepstack and layer_idx in self._deepstack_set:
|
| 186 |
+
# Apply the final layer_norm snapshot at this point, per MOSS's output_deepstack_hidden_states
|
| 187 |
+
# Actually, MOSS captures x BEFORE the final layer_norm — matches what PyTorch does.
|
| 188 |
+
deepstack.append(x)
|
| 189 |
+
|
| 190 |
+
x = self.layer_norm(x)
|
| 191 |
+
return x, (deepstack if return_deepstack else None)
|
| 192 |
+
|
| 193 |
+
|
| 194 |
+
# ---- GatedMLP (for audio_adapter + deepstack_audio_merger_list) ----
|
| 195 |
+
|
| 196 |
+
|
| 197 |
+
class GatedMLP(nn.Module):
|
| 198 |
+
"""MOSS's GatedMLP: down(silu(gate(x)) * up(x)). SwiGLU convention.
|
| 199 |
+
|
| 200 |
+
Matches MOSS/src/modeling_moss_audio.py:155-164.
|
| 201 |
+
All linears are bias=False.
|
| 202 |
+
"""
|
| 203 |
+
|
| 204 |
+
def __init__(self, input_size: int, hidden_size: int, output_size: int):
|
| 205 |
+
super().__init__()
|
| 206 |
+
self.gate_proj = nn.Linear(input_size, hidden_size, bias=False)
|
| 207 |
+
self.up_proj = nn.Linear(input_size, hidden_size, bias=False)
|
| 208 |
+
self.down_proj = nn.Linear(hidden_size, output_size, bias=False)
|
| 209 |
+
|
| 210 |
+
def __call__(self, x: mx.array) -> mx.array:
|
| 211 |
+
return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
|
| 212 |
+
|
| 213 |
+
|
| 214 |
+
__all__ = ["sinusoids", "WhisperAttention", "WhisperEncoderBlock",
|
| 215 |
+
"EncoderConfig", "MossAudioEncoderMLX", "GatedMLP"]
|
scripts/moss_audio_mel_mlx.py
ADDED
|
@@ -0,0 +1,192 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Pure-MLX mel-spectrogram + input_ids builder for MOSS-Audio.
|
| 2 |
+
|
| 3 |
+
Replaces the torch-dependent MossAudioProcessor at inference time. Matches the
|
| 4 |
+
upstream processor's output byte-for-byte on our test clips (see parity check
|
| 5 |
+
in test_moss_audio_mel_parity.py).
|
| 6 |
+
|
| 7 |
+
What's ported:
|
| 8 |
+
- Log-mel spectrogram (n_mels=128, n_fft=400, hop=160, sr=16000) via mx.fft.rfft,
|
| 9 |
+
reusing mlx-examples/whisper's mel_filters.npz for the filterbank.
|
| 10 |
+
- Whisper-style fbank normalization: log10 → clip to (max - 8.0) → (+4)/4.
|
| 11 |
+
- input_ids construction: audio-span expansion with 2-second time markers,
|
| 12 |
+
chat-template wrapping (<|im_start|>system/user/assistant).
|
| 13 |
+
|
| 14 |
+
What still uses a non-torch dep:
|
| 15 |
+
- Tokenization: HuggingFace AutoTokenizer (pure Python, no torch at runtime).
|
| 16 |
+
Install path: `pip install transformers` gets the slow BPE tokenizer; no
|
| 17 |
+
torch required at import.
|
| 18 |
+
"""
|
| 19 |
+
from __future__ import annotations
|
| 20 |
+
|
| 21 |
+
from functools import lru_cache
|
| 22 |
+
from pathlib import Path
|
| 23 |
+
from typing import Sequence
|
| 24 |
+
|
| 25 |
+
import mlx.core as mx
|
| 26 |
+
import numpy as np
|
| 27 |
+
|
| 28 |
+
SAMPLE_RATE = 16_000
|
| 29 |
+
N_FFT = 400
|
| 30 |
+
HOP_LENGTH = 160
|
| 31 |
+
MEL_DIM = 128
|
| 32 |
+
AUDIO_TOKENS_PER_SECOND = 12.5
|
| 33 |
+
|
| 34 |
+
AUDIO_TOKEN_ID = 151654
|
| 35 |
+
AUDIO_START_ID = 151669
|
| 36 |
+
AUDIO_END_ID = 151670
|
| 37 |
+
|
| 38 |
+
DIGIT_TOKEN_IDS = {str(i): 15 + i for i in range(10)}
|
| 39 |
+
|
| 40 |
+
_ASSETS = Path(__file__).resolve().parent / "assets"
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
@lru_cache(maxsize=None)
|
| 44 |
+
def _mel_filters() -> mx.array:
|
| 45 |
+
return mx.load(str(_ASSETS / "mel_filters.npz"))[f"mel_{MEL_DIM}"]
|
| 46 |
+
|
| 47 |
+
|
| 48 |
+
@lru_cache(maxsize=None)
|
| 49 |
+
def _hann_window(size: int) -> mx.array:
|
| 50 |
+
return mx.array(np.hanning(size + 1)[:-1])
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def _stft_rfft(x: mx.array) -> mx.array:
|
| 54 |
+
"""STFT with reflect padding, returning rfft bins. Matches librosa defaults."""
|
| 55 |
+
padding = N_FFT // 2
|
| 56 |
+
prefix = x[1 : padding + 1][::-1]
|
| 57 |
+
suffix = x[-(padding + 1) : -1][::-1]
|
| 58 |
+
x = mx.concatenate([prefix, x, suffix])
|
| 59 |
+
|
| 60 |
+
noverlap = HOP_LENGTH
|
| 61 |
+
t = (x.size - N_FFT + noverlap) // noverlap
|
| 62 |
+
strides = [noverlap, 1]
|
| 63 |
+
x = mx.as_strided(x, shape=[t, N_FFT], strides=strides)
|
| 64 |
+
return mx.fft.rfft(x * _hann_window(N_FFT))
|
| 65 |
+
|
| 66 |
+
|
| 67 |
+
def log_mel_spectrogram_mlx(audio: np.ndarray) -> mx.array:
|
| 68 |
+
"""Compute (MEL_DIM, n_frames) log-mel spectrogram in pure MLX.
|
| 69 |
+
|
| 70 |
+
Matches upstream's WhisperFeatureExtractor._np_extract_fbank_features with
|
| 71 |
+
n_mels=128. Used by the MOSS-Audio encoder, which crops/pads along the
|
| 72 |
+
frame axis.
|
| 73 |
+
"""
|
| 74 |
+
if not isinstance(audio, mx.array):
|
| 75 |
+
audio = mx.array(audio.astype(np.float32))
|
| 76 |
+
|
| 77 |
+
freqs = _stft_rfft(audio)
|
| 78 |
+
magnitudes = freqs[:-1, :].abs().square()
|
| 79 |
+
|
| 80 |
+
filters = _mel_filters()
|
| 81 |
+
mel_spec = magnitudes @ filters.T
|
| 82 |
+
log_spec = mx.maximum(mel_spec, 1e-10).log10()
|
| 83 |
+
log_spec = mx.maximum(log_spec, log_spec.max() - 8.0)
|
| 84 |
+
log_spec = (log_spec + 4.0) / 4.0
|
| 85 |
+
return log_spec.T # (MEL_DIM, n_frames)
|
| 86 |
+
|
| 87 |
+
|
| 88 |
+
def _conv3_downsample_len(raw_mel_len: int) -> int:
|
| 89 |
+
"""MOSS-Audio encoder stem: three conv/stride-2 layers → audio token count."""
|
| 90 |
+
n = int(raw_mel_len)
|
| 91 |
+
for _ in range(3):
|
| 92 |
+
n = (n - 1) // 2 + 1
|
| 93 |
+
return n
|
| 94 |
+
|
| 95 |
+
|
| 96 |
+
def _digit_token_ids(second: int) -> list[int]:
|
| 97 |
+
return [DIGIT_TOKEN_IDS[d] for d in str(second)]
|
| 98 |
+
|
| 99 |
+
|
| 100 |
+
def _build_audio_placeholder_ids(num_audio_tokens: int, *, enable_time_marker: bool) -> list[int]:
|
| 101 |
+
if not enable_time_marker:
|
| 102 |
+
return [AUDIO_TOKEN_ID] * num_audio_tokens
|
| 103 |
+
|
| 104 |
+
tokens_per_marker = int(AUDIO_TOKENS_PER_SECOND * 2) # every 2 seconds
|
| 105 |
+
total_seconds = num_audio_tokens / AUDIO_TOKENS_PER_SECOND
|
| 106 |
+
num_full_seconds = int(total_seconds)
|
| 107 |
+
|
| 108 |
+
out: list[int] = []
|
| 109 |
+
consumed = 0
|
| 110 |
+
for second in range(2, num_full_seconds + 1, 2):
|
| 111 |
+
marker_pos = (second // 2) * tokens_per_marker
|
| 112 |
+
segment_len = marker_pos - consumed
|
| 113 |
+
if segment_len > 0:
|
| 114 |
+
out.extend([AUDIO_TOKEN_ID] * segment_len)
|
| 115 |
+
consumed += segment_len
|
| 116 |
+
out.extend(_digit_token_ids(second))
|
| 117 |
+
|
| 118 |
+
remaining = num_audio_tokens - consumed
|
| 119 |
+
if remaining > 0:
|
| 120 |
+
out.extend([AUDIO_TOKEN_ID] * remaining)
|
| 121 |
+
return out
|
| 122 |
+
|
| 123 |
+
|
| 124 |
+
def _default_prompt(text: str) -> str:
|
| 125 |
+
return (
|
| 126 |
+
"<|im_start|>system\n"
|
| 127 |
+
"You are a helpful assistant.<|im_end|>\n"
|
| 128 |
+
"<|im_start|>user\n"
|
| 129 |
+
"<|audio_bos|><|AUDIO|><|audio_eos|>\n"
|
| 130 |
+
f"{text}<|im_end|>\n"
|
| 131 |
+
"<|im_start|>assistant\n"
|
| 132 |
+
)
|
| 133 |
+
|
| 134 |
+
|
| 135 |
+
def build_input_ids(tokenizer, prompt_text: str, *, audio_token_count: int,
|
| 136 |
+
enable_time_marker: bool = True) -> list[int]:
|
| 137 |
+
"""Expand the <|audio_bos|><|AUDIO|><|audio_eos|> marker into audio_token_count tokens.
|
| 138 |
+
|
| 139 |
+
Mirrors MossAudioProcessor._build_input_from_prompt but without torch tensors.
|
| 140 |
+
"""
|
| 141 |
+
import re
|
| 142 |
+
audio_span_re = re.compile(r"<\|audio_bos\|>(?:<\|AUDIO\|>)+<\|audio_eos\|>")
|
| 143 |
+
|
| 144 |
+
if audio_span_re.search(prompt_text) is None:
|
| 145 |
+
prompt_text = _default_prompt(prompt_text)
|
| 146 |
+
|
| 147 |
+
spans = list(audio_span_re.finditer(prompt_text))
|
| 148 |
+
if len(spans) != 1:
|
| 149 |
+
raise ValueError(f"Expected exactly 1 audio span, got {len(spans)}")
|
| 150 |
+
|
| 151 |
+
match = spans[0]
|
| 152 |
+
prefix = prompt_text[: match.start()]
|
| 153 |
+
suffix = prompt_text[match.end():]
|
| 154 |
+
|
| 155 |
+
ids: list[int] = []
|
| 156 |
+
if prefix:
|
| 157 |
+
ids.extend(tokenizer.encode(prefix, add_special_tokens=False))
|
| 158 |
+
ids.append(AUDIO_START_ID)
|
| 159 |
+
ids.extend(_build_audio_placeholder_ids(audio_token_count, enable_time_marker=enable_time_marker))
|
| 160 |
+
ids.append(AUDIO_END_ID)
|
| 161 |
+
if suffix:
|
| 162 |
+
ids.extend(tokenizer.encode(suffix, add_special_tokens=False))
|
| 163 |
+
return ids
|
| 164 |
+
|
| 165 |
+
|
| 166 |
+
def build_mel_and_input_ids(
|
| 167 |
+
audio: np.ndarray,
|
| 168 |
+
tokenizer,
|
| 169 |
+
*,
|
| 170 |
+
prompt: str,
|
| 171 |
+
enable_time_marker: bool = True,
|
| 172 |
+
) -> tuple[mx.array, mx.array, mx.array, int]:
|
| 173 |
+
"""Pure-MLX equivalent of MossAudioProcessor(text=prompt, audios=[audio]).
|
| 174 |
+
|
| 175 |
+
Returns (mel, lens, input_ids, audio_token_id) — same shapes and semantics
|
| 176 |
+
as the torch bridge's build_mel_spectrogram().
|
| 177 |
+
"""
|
| 178 |
+
audio_f32 = audio.astype(np.float32)
|
| 179 |
+
mel = log_mel_spectrogram_mlx(audio_f32) # (MEL_DIM, T)
|
| 180 |
+
raw_mel_len = int(mel.shape[-1])
|
| 181 |
+
audio_token_count = _conv3_downsample_len(raw_mel_len)
|
| 182 |
+
|
| 183 |
+
mel = mel[None, ...] # (1, MEL_DIM, T)
|
| 184 |
+
lens = mx.array(np.array([raw_mel_len], dtype=np.int32))
|
| 185 |
+
|
| 186 |
+
input_ids = build_input_ids(
|
| 187 |
+
tokenizer, prompt, audio_token_count=audio_token_count,
|
| 188 |
+
enable_time_marker=enable_time_marker,
|
| 189 |
+
)
|
| 190 |
+
input_ids_mx = mx.array(np.array([input_ids], dtype=np.int64))
|
| 191 |
+
|
| 192 |
+
return mel, lens, input_ids_mx, AUDIO_TOKEN_ID
|
scripts/moss_audio_mlx_bridge_v3.py
ADDED
|
@@ -0,0 +1,314 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""Pure-MLX MOSS-Audio hybrid bridge (v3). No PyTorch at runtime.
|
| 2 |
+
|
| 3 |
+
Uses:
|
| 4 |
+
- MLX INT4 Qwen3 LLM (existing moss4b_mlx_int4/)
|
| 5 |
+
- MLX BF16 MossAudioEncoder (ported to mlx.nn)
|
| 6 |
+
- MLX BF16 GatedMLP for audio_adapter + deepstack_audio_merger_list
|
| 7 |
+
- MossAudioProcessor from upstream repo (CPU/numpy only, used to compute mel spectrogram)
|
| 8 |
+
|
| 9 |
+
All compute on MLX; only mel-spectrogram construction uses CPU.
|
| 10 |
+
"""
|
| 11 |
+
from __future__ import annotations
|
| 12 |
+
|
| 13 |
+
import argparse
|
| 14 |
+
import sys
|
| 15 |
+
import time
|
| 16 |
+
from pathlib import Path
|
| 17 |
+
|
| 18 |
+
sys.path.insert(0, str(Path.home() / "benchmark" / "moss-audio"))
|
| 19 |
+
sys.path.insert(0, str(Path.home() / "benchmark" / "scripts"))
|
| 20 |
+
|
| 21 |
+
import librosa
|
| 22 |
+
import mlx.core as mx
|
| 23 |
+
import mlx.nn as mnn
|
| 24 |
+
import numpy as np
|
| 25 |
+
from mlx_lm import load as mlx_load
|
| 26 |
+
from mlx_lm.generate import generate_step
|
| 27 |
+
from mlx_lm.sample_utils import make_sampler, make_logits_processors
|
| 28 |
+
|
| 29 |
+
from moss_audio_encoder_mlx import MossAudioEncoderMLX, EncoderConfig, GatedMLP
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
# ---- Load MLX encoder + adapter + mergers ----
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def _infer_llm_hidden(ad_w: dict) -> int:
|
| 36 |
+
"""Sniff the LLM hidden dim from the adapter's down_proj.
|
| 37 |
+
|
| 38 |
+
Supports both BF16 weights (`down_proj.weight` shape = (llm_hidden, 8192))
|
| 39 |
+
and INT4-quantized weights (`down_proj.scales` shape = (llm_hidden, n_groups)).
|
| 40 |
+
Returns 2560 for 4B, 4096 for 8B.
|
| 41 |
+
"""
|
| 42 |
+
if "down_proj.scales" in ad_w:
|
| 43 |
+
return ad_w["down_proj.scales"].shape[0]
|
| 44 |
+
return ad_w["down_proj.weight"].shape[0]
|
| 45 |
+
|
| 46 |
+
|
| 47 |
+
def load_mlx_audio_path(weights_dir: Path, *, int4: bool = False):
|
| 48 |
+
"""Load encoder + adapter + mergers into MLX.
|
| 49 |
+
|
| 50 |
+
Adapter/merger output dim (LLM hidden) is inferred from the saved weights,
|
| 51 |
+
so the same loader works for both 4B (2560) and 8B (4096).
|
| 52 |
+
|
| 53 |
+
When int4=True, weights_dir is expected to contain already-quantized
|
| 54 |
+
INT4 safetensors (see scripts/save_moss_audio_int4.py). This avoids
|
| 55 |
+
the double-buffer memory cost of live quantization from BF16.
|
| 56 |
+
"""
|
| 57 |
+
ad_w = mx.load(str(weights_dir / "audio_adapter.safetensors"))
|
| 58 |
+
llm_hidden = _infer_llm_hidden(ad_w)
|
| 59 |
+
|
| 60 |
+
cfg = EncoderConfig()
|
| 61 |
+
enc = MossAudioEncoderMLX(cfg)
|
| 62 |
+
adapter = GatedMLP(1280, 8192, llm_hidden)
|
| 63 |
+
mergers = [GatedMLP(1280, 8192, llm_hidden) for _ in range(3)]
|
| 64 |
+
|
| 65 |
+
if int4:
|
| 66 |
+
# Build quantized module structure FIRST, then load INT4 weights directly.
|
| 67 |
+
# This avoids holding BF16 + INT4 copies in memory simultaneously.
|
| 68 |
+
mnn.quantize(enc, group_size=64, bits=4)
|
| 69 |
+
mnn.quantize(adapter, group_size=64, bits=4)
|
| 70 |
+
for m in mergers:
|
| 71 |
+
mnn.quantize(m, group_size=64, bits=4)
|
| 72 |
+
|
| 73 |
+
enc_w = mx.load(str(weights_dir / "audio_encoder.safetensors"))
|
| 74 |
+
enc.load_weights(list(enc_w.items()), strict=True)
|
| 75 |
+
|
| 76 |
+
adapter.load_weights(list(ad_w.items()), strict=True)
|
| 77 |
+
|
| 78 |
+
dm_w = mx.load(str(weights_dir / "deepstack_mergers.safetensors"))
|
| 79 |
+
for i, merger in enumerate(mergers):
|
| 80 |
+
subw = {k[len(f"{i}."):]: v for k, v in dm_w.items() if k.startswith(f"{i}.")}
|
| 81 |
+
merger.load_weights(list(subw.items()), strict=True)
|
| 82 |
+
|
| 83 |
+
mx.eval(enc.parameters(), adapter.parameters(), *[m.parameters() for m in mergers])
|
| 84 |
+
return enc, adapter, mergers
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
def build_mel_spectrogram(audio_np: np.ndarray, processor_or_tokenizer, *, fp32_compute: bool = True) -> tuple[mx.array, mx.array]:
|
| 88 |
+
"""Build mel spectrogram + audio-expanded input_ids.
|
| 89 |
+
|
| 90 |
+
Two supported inputs for `processor_or_tokenizer`:
|
| 91 |
+
- A HuggingFace tokenizer (pure-Python, no torch) — pure-MLX path. Default.
|
| 92 |
+
- A MossAudioProcessor instance — delegates to upstream (kept for the
|
| 93 |
+
parity regression tests and anyone pinning to the old behavior).
|
| 94 |
+
|
| 95 |
+
`fp32_compute=True` (default) keeps the mel input as fp32, which causes
|
| 96 |
+
MLX to auto-promote all encoder activations to fp32 during forward
|
| 97 |
+
(weights stay BF16 on disk). This cuts relative error in the adapter
|
| 98 |
+
output from 6.5% to 0.98% vs PyTorch reference, at the cost of ~1 GB
|
| 99 |
+
extra peak memory during the one-shot encoder pass.
|
| 100 |
+
"""
|
| 101 |
+
prompt = ("Describe this audio in detail. Include speech content, speaker "
|
| 102 |
+
"characteristics, background sounds, music, and any notable temporal events.")
|
| 103 |
+
|
| 104 |
+
# Pure-MLX path: the object has `.encode()` but not MossAudioProcessor's `audio_token_id`
|
| 105 |
+
# set alongside a `.tokenizer` attribute. Detect by method presence.
|
| 106 |
+
is_tokenizer = hasattr(processor_or_tokenizer, "encode") and not hasattr(processor_or_tokenizer, "audio_token_id")
|
| 107 |
+
|
| 108 |
+
if is_tokenizer:
|
| 109 |
+
from moss_audio_mel_mlx import build_mel_and_input_ids
|
| 110 |
+
mel_mx, lens_mx, input_ids_mx, audio_token_id = build_mel_and_input_ids(
|
| 111 |
+
audio_np, processor_or_tokenizer, prompt=prompt, enable_time_marker=True,
|
| 112 |
+
)
|
| 113 |
+
if not fp32_compute:
|
| 114 |
+
mel_mx = mel_mx.astype(mx.bfloat16)
|
| 115 |
+
return mel_mx, lens_mx, input_ids_mx, audio_token_id
|
| 116 |
+
|
| 117 |
+
# Legacy torch processor path (unchanged)
|
| 118 |
+
processor = processor_or_tokenizer
|
| 119 |
+
inputs = processor(text=prompt, audios=[audio_np], return_tensors="pt")
|
| 120 |
+
import torch
|
| 121 |
+
mel = inputs["audio_data"]
|
| 122 |
+
mel_np = mel.to(torch.float32).numpy()
|
| 123 |
+
mel_mx = mx.array(mel_np) if fp32_compute else mx.array(mel_np).astype(mx.bfloat16)
|
| 124 |
+
lens = inputs["audio_data_seqlens"].cpu().numpy().astype(np.int32) if inputs.get("audio_data_seqlens") is not None else None
|
| 125 |
+
lens_mx = mx.array(lens) if lens is not None else None
|
| 126 |
+
input_ids_np = inputs["input_ids"].cpu().numpy()
|
| 127 |
+
input_ids_mx = mx.array(input_ids_np)
|
| 128 |
+
audio_token_id = processor.audio_token_id
|
| 129 |
+
return mel_mx, lens_mx, input_ids_mx, audio_token_id
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
def run_mlx_audio_pipeline(encoder, adapter, mergers,
|
| 133 |
+
mel: mx.array, lens: mx.array):
|
| 134 |
+
"""Returns (primary_embeds, deepstack_embeds) all on MLX."""
|
| 135 |
+
last, deepstack = encoder(mel, feature_lens=lens, return_deepstack=True)
|
| 136 |
+
primary = adapter(last) # (B, N_audio, llm_hidden)
|
| 137 |
+
ds_embeds = [mergers[i](ds) for i, ds in enumerate(deepstack)]
|
| 138 |
+
return primary, ds_embeds
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
# ---- DeepStack injection on MLX decoder ----
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
def install_deepstack_hooks(mlx_model, deepstack_embeds: list[mx.array], audio_positions: np.ndarray):
|
| 145 |
+
"""Same class-level Qwen3Model.__call__ override as v2, but with MLX-native inputs."""
|
| 146 |
+
from mlx_lm.models.base import create_attention_mask
|
| 147 |
+
|
| 148 |
+
audio_positions_mx = mx.array(audio_positions.astype(np.int32))
|
| 149 |
+
num_inject = len(deepstack_embeds)
|
| 150 |
+
|
| 151 |
+
ModelCls = type(mlx_model.model)
|
| 152 |
+
orig_call = ModelCls.__call__
|
| 153 |
+
|
| 154 |
+
def new_call(self, inputs, cache=None, input_embeddings=None):
|
| 155 |
+
if self is not mlx_model.model:
|
| 156 |
+
return orig_call(self, inputs, cache, input_embeddings)
|
| 157 |
+
|
| 158 |
+
if input_embeddings is not None:
|
| 159 |
+
h = input_embeddings
|
| 160 |
+
else:
|
| 161 |
+
h = self.embed_tokens(inputs)
|
| 162 |
+
|
| 163 |
+
if cache is None:
|
| 164 |
+
cache = [None] * len(self.layers)
|
| 165 |
+
mask = create_attention_mask(h, cache[0])
|
| 166 |
+
is_prefill = h.shape[1] > 1
|
| 167 |
+
|
| 168 |
+
for layer_idx, (layer, c) in enumerate(zip(self.layers, cache)):
|
| 169 |
+
h = layer(h, mask, c)
|
| 170 |
+
if is_prefill and layer_idx < num_inject:
|
| 171 |
+
ds = deepstack_embeds[layer_idx]
|
| 172 |
+
if ds.dtype != h.dtype:
|
| 173 |
+
ds = ds.astype(h.dtype)
|
| 174 |
+
if ds.ndim == 3:
|
| 175 |
+
ds = ds[0] # flatten batch
|
| 176 |
+
h_batch0 = h[0]
|
| 177 |
+
h_batch0 = h_batch0.at[audio_positions_mx].add(ds)
|
| 178 |
+
h = h_batch0[None]
|
| 179 |
+
|
| 180 |
+
return self.norm(h)
|
| 181 |
+
|
| 182 |
+
ModelCls.__call__ = new_call
|
| 183 |
+
|
| 184 |
+
|
| 185 |
+
# ---- Main ----
|
| 186 |
+
|
| 187 |
+
|
| 188 |
+
def main():
|
| 189 |
+
parser = argparse.ArgumentParser()
|
| 190 |
+
parser.add_argument("--mlx-llm", default=str(Path.home() / "benchmark" / "moss4b_mlx_int4"))
|
| 191 |
+
parser.add_argument("--mlx-audio", default=str(Path.home() / "benchmark" / "moss4b_audio_mlx_int4"))
|
| 192 |
+
parser.add_argument("--moss-source", default="OpenMOSS-Team/MOSS-Audio-4B-Thinking",
|
| 193 |
+
help="HF repo for the processor only (mel spectrogram computation)")
|
| 194 |
+
parser.add_argument("--audio", required=True)
|
| 195 |
+
parser.add_argument("--max-tokens", type=int, default=2048)
|
| 196 |
+
parser.add_argument("--temp", type=float, default=1.0)
|
| 197 |
+
parser.add_argument("--int4-audio", action="store_true", default=True,
|
| 198 |
+
help="Load pre-quantized INT4 audio encoder + adapter + mergers from --mlx-audio")
|
| 199 |
+
parser.add_argument("--no-int4-audio", dest="int4_audio", action="store_false",
|
| 200 |
+
help="Use BF16 audio weights instead of INT4")
|
| 201 |
+
parser.add_argument("--repetition-penalty", type=float, default=1.02,
|
| 202 |
+
help="mlx_lm repetition_penalty. 1.02 is the shipped EN-scope default "
|
| 203 |
+
"(kills loop-decode on non-speech clips without starving genre descriptions).")
|
| 204 |
+
parser.add_argument("--repetition-context-size", type=int, default=20)
|
| 205 |
+
parser.add_argument("--use-torch-processor", action="store_true",
|
| 206 |
+
help="Use upstream MossAudioProcessor (torch+torchaudio) for mel "
|
| 207 |
+
"computation. Default off: pure-MLX mel path has <0.2%% "
|
| 208 |
+
"rel-err parity and no torch dep.")
|
| 209 |
+
args = parser.parse_args()
|
| 210 |
+
|
| 211 |
+
print(f"[mlx] loading LLM from {args.mlx_llm}", flush=True)
|
| 212 |
+
t0 = time.perf_counter()
|
| 213 |
+
mlx_model, mlx_tokenizer = mlx_load(args.mlx_llm)
|
| 214 |
+
print(f"[mlx] LLM loaded in {time.perf_counter()-t0:.1f}s peak={mx.get_peak_memory()/1e9:.2f}GB", flush=True)
|
| 215 |
+
|
| 216 |
+
print(f"[mlx] loading audio path from {args.mlx_audio} (int4={args.int4_audio})", flush=True)
|
| 217 |
+
t0 = time.perf_counter()
|
| 218 |
+
encoder, adapter, mergers = load_mlx_audio_path(Path(args.mlx_audio), int4=args.int4_audio)
|
| 219 |
+
print(f"[mlx] audio path loaded in {time.perf_counter()-t0:.1f}s peak={mx.get_peak_memory()/1e9:.2f}GB", flush=True)
|
| 220 |
+
|
| 221 |
+
# Pure-MLX path: use the MLX-LLM tokenizer directly for text encoding.
|
| 222 |
+
# Mel spectrogram + input_ids expansion are computed in pure MLX (no torch).
|
| 223 |
+
# The upstream MossAudioProcessor is only needed if you opt into the legacy
|
| 224 |
+
# torch mel path via `--use-torch-processor` (not on by default).
|
| 225 |
+
processor = mlx_tokenizer
|
| 226 |
+
if args.use_torch_processor:
|
| 227 |
+
from src.processing_moss_audio import MossAudioProcessor
|
| 228 |
+
processor = MossAudioProcessor.from_pretrained(
|
| 229 |
+
args.moss_source, trust_remote_code=True, enable_time_marker=True,
|
| 230 |
+
)
|
| 231 |
+
|
| 232 |
+
y, _ = librosa.load(args.audio, sr=16000, mono=True)
|
| 233 |
+
y = y.astype(np.float32)
|
| 234 |
+
print(f"[audio] {args.audio}: {len(y)/16000:.1f}s", flush=True)
|
| 235 |
+
|
| 236 |
+
t0 = time.perf_counter()
|
| 237 |
+
mel, lens, input_ids_mx, audio_token_id = build_mel_spectrogram(y, processor)
|
| 238 |
+
primary, ds_embeds = run_mlx_audio_pipeline(encoder, adapter, mergers, mel, lens)
|
| 239 |
+
# Cast to bf16 and force materialization NOW, so the fp32 activations can be freed
|
| 240 |
+
# before we start the decode phase (the encoder's fp32 forward is our peak-memory hotspot).
|
| 241 |
+
primary = primary.astype(mx.bfloat16)
|
| 242 |
+
ds_embeds = [d.astype(mx.bfloat16) for d in ds_embeds]
|
| 243 |
+
mx.eval(primary, *ds_embeds)
|
| 244 |
+
print(f"[mlx] audio encoded in {time.perf_counter()-t0:.2f}s primary={primary.shape}", flush=True)
|
| 245 |
+
print(f"[mem] after encode (pre-cleanup): peak={mx.get_peak_memory()/1e9:.2f}GB active={mx.get_active_memory()/1e9:.2f}GB", flush=True)
|
| 246 |
+
|
| 247 |
+
# FREE ENCODER + ADAPTER + MERGERS now that we have the embeddings.
|
| 248 |
+
# These modules are only needed at prefill; all downstream work (merge,
|
| 249 |
+
# deepstack injection, decode) uses only primary + ds_embeds tensors.
|
| 250 |
+
# Dropping them reclaims ~1.3 GB for clip_01 (up to ~2.5 GB if fp32
|
| 251 |
+
# activations were still held). Captured closures for install_deepstack_hooks
|
| 252 |
+
# keep only the embedding mx.array references, not the modules.
|
| 253 |
+
del encoder, adapter, mergers, mel, lens
|
| 254 |
+
import gc; gc.collect()
|
| 255 |
+
mx.clear_cache()
|
| 256 |
+
try: mx.reset_peak_memory()
|
| 257 |
+
except Exception: pass
|
| 258 |
+
print(f"[mem] after free encoder: peak={mx.get_peak_memory()/1e9:.2f}GB active={mx.get_active_memory()/1e9:.2f}GB", flush=True)
|
| 259 |
+
|
| 260 |
+
# Build merged text+audio embeddings
|
| 261 |
+
audio_mask = input_ids_mx == audio_token_id
|
| 262 |
+
audio_positions = np.where(np.array(audio_mask[0]))[0]
|
| 263 |
+
assert len(audio_positions) == primary.shape[1], \
|
| 264 |
+
f"mask positions {len(audio_positions)} != primary len {primary.shape[1]}"
|
| 265 |
+
|
| 266 |
+
text_embeds = mlx_model.model.embed_tokens(input_ids_mx) # (1, seq, hidden)
|
| 267 |
+
text_np = np.array(text_embeds.astype(mx.float32))
|
| 268 |
+
primary_np = np.array(primary.astype(mx.float32))
|
| 269 |
+
text_np[0, audio_positions, :] = primary_np[0, :, :]
|
| 270 |
+
merged = mx.array(text_np).astype(mx.bfloat16)
|
| 271 |
+
print(f"[bridge] merged embeds {merged.shape}", flush=True)
|
| 272 |
+
|
| 273 |
+
# Install DeepStack
|
| 274 |
+
ds_flat = [d[0] for d in ds_embeds] # drop batch dim
|
| 275 |
+
install_deepstack_hooks(mlx_model, ds_flat, audio_positions)
|
| 276 |
+
print(f"[deepstack] installed {len(ds_flat)} layer injections", flush=True)
|
| 277 |
+
|
| 278 |
+
# Generate
|
| 279 |
+
print(f"[gen] decoding max_tokens={args.max_tokens}...", flush=True)
|
| 280 |
+
sampler = make_sampler(temp=args.temp, top_p=1.0, top_k=50)
|
| 281 |
+
gen_kwargs = dict(
|
| 282 |
+
prompt=input_ids_mx[0],
|
| 283 |
+
model=mlx_model,
|
| 284 |
+
input_embeddings=merged[0],
|
| 285 |
+
max_tokens=args.max_tokens,
|
| 286 |
+
sampler=sampler,
|
| 287 |
+
)
|
| 288 |
+
if args.repetition_penalty:
|
| 289 |
+
gen_kwargs["logits_processors"] = make_logits_processors(
|
| 290 |
+
repetition_penalty=args.repetition_penalty,
|
| 291 |
+
repetition_context_size=args.repetition_context_size,
|
| 292 |
+
)
|
| 293 |
+
print(f"[gen] repetition_penalty={args.repetition_penalty} ctx={args.repetition_context_size}", flush=True)
|
| 294 |
+
t0 = time.perf_counter()
|
| 295 |
+
generated = []
|
| 296 |
+
ttft = None
|
| 297 |
+
for tok, _ in generate_step(**gen_kwargs):
|
| 298 |
+
if ttft is None:
|
| 299 |
+
ttft = time.perf_counter() - t0
|
| 300 |
+
generated.append(int(tok))
|
| 301 |
+
if tok == mlx_tokenizer.eos_token_id:
|
| 302 |
+
break
|
| 303 |
+
elapsed = time.perf_counter() - t0
|
| 304 |
+
n_tok = len(generated)
|
| 305 |
+
decode_s = max(elapsed - (ttft or 0), 1e-6)
|
| 306 |
+
print(f"[gen] elapsed={elapsed:.2f}s ttft={ttft:.3f}s out={n_tok}tok decode={max(n_tok-1,0)/decode_s:.1f}t/s", flush=True)
|
| 307 |
+
text = mlx_tokenizer.decode(generated)
|
| 308 |
+
print(f"\n=== OUTPUT ===\n{text[:1500]}\n=== END ===\n")
|
| 309 |
+
|
| 310 |
+
print(f"[mem] mlx peak={mx.get_peak_memory()/1e9:.2f}GB", flush=True)
|
| 311 |
+
|
| 312 |
+
|
| 313 |
+
if __name__ == "__main__":
|
| 314 |
+
main()
|