Qwen3.5-122B-A10B · NVFP4A16 (W4A16, FP4 weights)
4-bit quantization of Qwen/Qwen3.5-122B-A10B, produced for efficient vLLM serving. Quantized by axteam.
Revision 2026-11-01 (rebuilt): The MoE router (
mlp.gate) and the shared-expert gate are now compensated during AWQ smoothing (checked on every layer against the original: the router rescale matches the norm rescale, median per-channel difference below 2%). Channels whose original norm gain is zero are now left out of AWQ's scale normalization, so the smoothed norms keep their BF16 precision (checked: in every smoothed norm, half a BF16 step relative to the gain is below 1%, median over channels). The first upload of this repo left the router and the shared-expert gate uncompensated, so tokens were routed to different experts than in the original, and lost norm precision in layer 27 because of a zero-gain channel (the stored gains there could only be represented to about ±5%: half a BF16 step relative to the gain, median over channels). If you downloaded this repo before 2026-11-01 11:26:43 UTC (revision10716cc), download it again. The benchmark and fidelity figures on this card were measured on the rebuilt files.
| Base model | Qwen/Qwen3.5-122B-A10B |
| Method | NVFP4A16 (W4A16, FP4 weights) |
| Format | compressed-tensors nvfp4-pack-quantized |
| Weights / activations | W4A16 — FP4 (E2M1) weights group 16 (FP8 E4M3 scale + per-tensor FP32 global scale), BF16 activations |
| Architecture | Hybrid (linear-attention + full-attention) · MoE (256 experts + shared, ~10B / 122B total active) · multimodal (vision) · MTP head |
| Size on disk | ~81.5 GB |
| Calibration | 256 samples (Korean-heavy mixed corpus) |
| License | Apache-2.0 |
Quantization details
- Scheme — W4A16 — FP4 (E2M1) weights group 16 (FP8 E4M3 scale + per-tensor FP32 global scale), BF16 activations, stored as compressed-tensors
nvfp4-pack-quantized. - Kept in BF16 — the vision tower & projector (
model.visual.*), the multi-token-prediction head (mtp.*), all RMSNorms, token embeddings, the LM head, the MoE router (mlp.gate) and shared-expert gate, the entire linear-attention (GatedDeltaNet/SSM) path — FP4 is too coarse for the SSM weights, so the fulllinear_attn.*stays BF16.linear_attn.A_logand the gatedlinear_attn.norm, FP32 in the original, are stored in BF16 — losslessly forA_log(every original value is exactly representable in BF16), and for the norm the same values vLLM and Transformers (default dtype) load at runtime. Left unquantized to preserve quality (about 6% of all parameters in this build). - Tokenizer /
chat_template/generation_configare identical to the original.config.jsonis the original plus aquantization_configblock.
This is post-training quantization (PTQ) — no retraining. AWQ uses a small calibration set (no gradient) to protect salient weight channels before rounding.
Benchmarks (measured)
Every column was measured with the same harness and settings — lm-evaluation-harness on vLLM, greedy, completion-style prompts without a chat template (0- or 5-shot, as listed in each row). MMLU here is lm-eval's generative variant (mmlu_generative: the model writes the answer letter and only an exact match on the first line counts), so it is not directly comparable with MMLU scores measured another way (e.g. log-likelihood scoring). Percentages are recovery vs. the BF16 original.
| Benchmark | BF16 (original) | FP8-dynamic | AWQ-8bit | NVFP4A16 (this repo) | NVFP4 | AWQ-4bit |
|---|---|---|---|---|---|---|
| MMLU (5-shot, generative) | 88.06 | 87.89 (99.8%) | 87.89 (99.8%) | 87.91 (99.8%) | 87.69 (99.6%) | 87.69 (99.6%) |
| GSM8K (5-shot, strict) | 84.46 | 83.93 (99.4%) | 84.23 (99.7%) | 84.84 (100.4%) | 80.52 (95.3%) | 83.78 (99.2%) |
| ARC-Challenge (0-shot) | 63.74 | 63.91 (100.3%) | 63.74 (100.0%) | 63.40 (99.5%) | 62.71 (98.4%) | 62.80 (98.5%) |
| Winogrande (5-shot) | 73.09 | 73.24 (100.2%) | 73.56 (100.6%) | 72.77 (99.6%) | 73.16 (100.1%) | 73.01 (99.9%) |
| KMMLU (5-shot, direct) | 71.72 | 71.72 (100.0%) | 71.67 (99.9%) | 71.35 (99.5%) | 70.80 (98.7%) | 71.20 (99.3%) |
| HAE-RAE (0-shot) | 82.31 | 81.39 (98.9%) | 81.76 (99.3%) | 81.48 (99.0%) | 80.20 (97.4%) | 80.93 (98.3%) |
| Average | 77.23 | 77.01 (99.7%) | 77.14 (99.9%) | 76.96 (99.7%) | 75.85 (98.2%) | 76.57 (99.1%) |
Evaluation (measured)
Fidelity to the BF16 original on an internal mixed English/Korean set (100 prompts: the BF16 original's greedy answers, scored under each build with the same prompts and scoring code). This build was measured on the rebuilt files with vLLM 0.32 on 2026-11-02; the BF16 reference column was measured on 2026-11-02.
| Metric | This model | BF16 original |
|---|---|---|
| Perplexity (lower = closer) | 1.3507 | 1.3258 (+1.9%) |
| Top-1 next-token agreement | 95.2% | 98.4% |
These measure fidelity to the base model, not absolute task scores.
Requirements
| Runtime | vLLM — the architecture is in vLLM ≥ 0.17.0; tested here with vLLM 0.32.0. |
| vLLM install | vLLM 0.32 itself needs Python 3.11–3.14, torch 2.13 (CUDA 13.0 build, so an NVIDIA driver with CUDA 13 support), transformers ≥5.16.1,<5.20 and compressed-tensors 0.17.0 (both pinned by vLLM; 0.17.0 reads this checkpoint, which was saved with 0.19.0 — do not upgrade it separately). |
| GPU | NVIDIA Turing (sm75) or newer — weight-only FP4, no FP4 tensor cores needed. V100 (sm70) is not supported. (Kernel choice per GPU from vLLM 0.32's selection code; only B200 was run here.) |
| Memory | Weights: 81.5 GB. Add KV cache (grows with context length × concurrent requests) and a few GB for activations / CUDA graphs; use --tensor-parallel-size N when the weights do not fit one GPU. |
| Transformers (without vLLM) | Not loadable correctly by Transformers — use vLLM. Up to 5.22.1 its MoE loader cannot merge per-expert 4-bit weights of this format (checked on the first 2 layers of the Qwen3.6-35B-A3B AWQ-4bit and NVFP4A16 builds; this repo itself was not loaded in Transformers). |
| trust_remote_code | not needed |
Usage (vLLM)
Start an OpenAI-compatible server:
vllm serve axteam/Qwen3.5-122B-A10B-NVFP4A16 --max-model-len 32768
Query it:
curl http://localhost:8000/v1/chat/completions -H "Content-Type: application/json" -d '{
"model": "axteam/Qwen3.5-122B-A10B-NVFP4A16",
"messages": [{"role": "user", "content": "대한민국의 수도는 어디인가요?"}],
"temperature": 0.6, "top_p": 0.95, "top_k": 20, "max_tokens": 512
}'
Offline inference in Python:
from vllm import LLM, SamplingParams
llm = LLM(model="axteam/Qwen3.5-122B-A10B-NVFP4A16", max_model_len=32768)
params = SamplingParams(temperature=0.6, top_p=0.95, top_k=20, max_tokens=1024)
out = llm.chat([{"role": "user", "content": "Explain model quantization in two sentences."}], params)
print(out[0].outputs[0].text)
# image input (URL or data: URI)
messages = [{"role": "user", "content": [
{"type": "image_url", "image_url": {"url": "https://example.com/your-image.png"}},
{"type": "text", "text": "Describe this image."},
]}]
out = llm.chat(messages, params)
print(out[0].outputs[0].text)
Recommended sampling (from the original
generation_config.json):temperature=0.6,top_p=0.95,top_k=20. Thinking mode is on by default in the chat template; to turn it off passchat_template_kwargs={"enable_thinking": False}(Pythonllm.chat(..., chat_template_kwargs=...), or"chat_template_kwargs": {"enable_thinking": false}in the request body). Text-only serving: add--limit-mm-per-prompt '{"image": 0, "video": 0}'to skip loading the vision encoder's memory budget.
Reducing KV-cache memory (optional, serve-time)
The KV cache is kept in BF16 by default. KV-cache precision is a runtime choice, independent of the weight quantization above. This is a hybrid model: only its 12 full-attention layers keep a KV cache, while its 36 linear-attention (Gated DeltaNet) layers keep a fixed-size state per request, which --kv-cache-dtype does not change. FP8 halves the full-attention KV cache, so the saving grows with context length and is small for many short requests:
vllm serve axteam/Qwen3.5-122B-A10B-NVFP4A16 --kv-cache-dtype fp8
License
Quantized derivative of Qwen/Qwen3.5-122B-A10B (Apache-2.0). All original weights, tokenizer and chat template belong to the Qwen team.
한국어 요약
Qwen/Qwen3.5-122B-A10B의 NVFP4A16 (W4A16) 양자화 빌드 · 제작 axteam
개요
| 항목 | 내용 |
|---|---|
| 원본 모델 | Qwen/Qwen3.5-122B-A10B |
| 양자화 방식 | NVFP4A16 (W4A16, FP4 가중치) |
| 저장 형식 | compressed-tensors nvfp4-pack-quantized |
| 모델 구조 | 하이브리드 어텐션(선형 어텐션 + 풀 어텐션) · MoE (전문가 256개 + 공유 전문가, 총 122B 중 약 10B 활성) · 멀티모달(비전) · MTP 헤드 |
| 용량 | 약 81.5 GB |
| 보정 데이터 | 샘플 256개 (한국어 위주 혼합 코퍼스) |
양자화 상세
- 가중치: FP4(E2M1), 그룹 크기 16 (그룹별 FP8(E4M3) 스케일 + 텐서별 FP32 전역 스케일)
- 활성화: BF16 (양자화하지 않음)
- 양자화 절차: PTQ(학습 후 양자화, 재학습 없음), AWQ 방식 — 보정 데이터로 활성화가 큰 가중치 채널을 찾아 채널별 스케일로 보호한 뒤 FP4로 반올림
- BF16 유지: 비전 타워·투영층(
model.visual.*), MTP 헤드(mtp.*), 모든 RMSNorm, 토큰 임베딩, LM 헤드, MoE 라우터(mlp.gate)와 공유 전문가 게이트, 선형 어텐션(GatedDeltaNet/SSM) 경로 전체(linear_attn.*, FP4로는 정밀도 부족) — 원본에서 FP32인linear_attn.A_log·게이트linear_attn.norm은 BF16으로 저장(A_log는 원본 값이 모두 BF16으로 정확히 표현돼 손실 없음, norm은 vLLM·Transformers(기본 dtype)가 실행 때 쓰는 BF16 값과 같음) — 전체 파라미터의 약 6% - 원본 파일: 토크나이저·chat template·generation config는 원본과 동일,
config.json은 원본에quantization_config블록만 추가
품질 (실측)
- 원본 대비 충실도: PPL 1.3507 (원본 1.3258, +1.9%, 낮을수록 원본에 가까움), 다음 토큰 top-1 일치율 95.2% (원본 자체 기준값 98.4%) — 내부 영어·한국어 혼합 프롬프트 100개, 원본(BF16)의 greedy 답변을 기준 텍스트로 측정
- 벤치마크 평균: 76.96 (원본 77.23, 99.7%) — MMLU(생성형)·GSM8K·ARC-Challenge·Winogrande·KMMLU·HAE-RAE 6종 평균, lm-evaluation-harness + vLLM, chat template 없는 완성형 프롬프트(과제별 0·5-shot, 표 참고)·greedy, 모든 빌드를 같은 설정으로 측정
- 한국어 벤치마크: KMMLU 71.35 (원본 71.72, 99.5%) · HAE-RAE 81.48 (원본 82.31, 99.0%)
- MMLU 방식: lm-eval 생성형 변형(
mmlu_generative) — 모델이 답 글자를 생성하고 첫 줄이 정확히 일치해야 정답으로 채점하므로, 다른 방식(예: log-likelihood 채점)으로 잰 MMLU 점수와 직접 비교 불가
요구사항
- 실행 엔진: vLLM 0.17.0 이상(이 구조를 지원하는 첫 버전), 이 빌드는 vLLM 0.32.0에서 실행 확인
- vLLM 0.32 설치 조건: Python 3.11~3.14, torch 2.13(CUDA 13.0 빌드, CUDA 13 지원 드라이버 필요), transformers 5.16.1 이상 5.20 미만, compressed-tensors 0.17.0 — transformers·compressed-tensors는 vLLM이 버전을 고정하므로 compressed-tensors를 따로 업그레이드하지 말 것 (이 체크포인트는 0.19.0으로 저장됐지만 0.17.0에서도 로드됨)
- GPU: NVIDIA Turing(sm75) 이상, V100(sm70) 미지원 — 가중치 전용 FP4라 FP4 전용 텐서코어 불필요 (GPU별 지원 여부는 vLLM 0.32 커널 선택 코드 기준, 실제 실행은 B200에서만 확인)
- 메모리: 가중치 81.5 GB + KV 캐시(컨텍스트 길이 × 동시 요청 수에 비례) + 활성화·CUDA graph용 수 GB — GPU 한 장에 다 들어가지 않으면
--tensor-parallel-size N으로 분산 - Transformers로 직접 사용: 불가 — 이 형식의 전문가별 4비트 가중치를 5.22.1까지의 MoE 로더가 합치지 못함 (Qwen3.6-35B-A3B AWQ-4bit·NVFP4A16 빌드의 앞 2개 층으로 확인, 이 리포 자체는 Transformers 로드 미실행), vLLM 사용
- trust_remote_code: 불필요
사용법
- 서버 실행:
vllm serve axteam/Qwen3.5-122B-A10B-NVFP4A16 --max-model-len 32768— OpenAI 호환 API - Python:
LLM(model="axteam/Qwen3.5-122B-A10B-NVFP4A16", max_model_len=32768).chat(...)— 전체 예시는 위 영어 본문 'Usage (vLLM)' 절 참고 - 권장 샘플링:
temperature=0.6,top_p=0.95,top_k=20(원본generation_config.json기준) - thinking 모드: 기본 켜짐 — 끄려면
chat_template_kwargs={"enable_thinking": False}전달(Python은llm.chat()인자, HTTP는 요청 본문의"chat_template_kwargs": {"enable_thinking": false}) - 이미지 입력: 메시지에
image_url항목 사용 - 텍스트만 사용할 때:
--limit-mm-per-prompt '{"image": 0, "video": 0}'— 비전 타워용 메모리 예약을 생략해 메모리 절약 - KV 캐시 절감(선택):
vllm serve에--kv-cache-dtype fp8추가 — 하이브리드 구조라 KV 캐시는 풀 어텐션 12개 층에만 있고, 선형 어텐션(Gated DeltaNet) 36개 층의 요청별 고정 크기 상태는 이 옵션으로 바뀌지 않음 → 풀 어텐션 KV 캐시만 절반으로 줄어 긴 컨텍스트에서 효과가 크고 짧은 요청이 많을 때는 작음 (가중치 양자화와 별개인 실행 옵션)
참고
- 수정 이력 (2026-11-01, 재양자화): MoE 라우터(
mlp.gate)·공유 전문가 게이트에 AWQ 상쇄 스케일 적용 (모든 층에서 원본 대비 라우터 배율이 norm 배율과 일치함을 확인, 채널별 차이 중앙값 2% 미만); 원본 norm 이득이 0인 채널을 AWQ 스케일 정규화에서 빼 스무딩된 norm의 BF16 정밀도 유지 (스무딩된 모든 norm에서 이득 대비 BF16 반 칸이 1% 미만임을 확인, 채널 중앙값) — 이전 업로드의 결함: 첫 번째 업로드는 라우터·공유 전문가 게이트 상쇄 누락으로 원본과 다른 전문가 선택, 27번 층의 이득 0 채널 탓에 norm 정밀도 손실(그 층 이득을 BF16으로 약 ±5%까지만 나타낼 수 있었음: 이득 대비 BF16 반 칸, 채널 중앙값) — 2026-11-01 11:26:43 UTC(리비전10716cc) 이전에 받았다면 다시 받을 것, 품질 항목의 벤치마크·충실도는 재양자화 빌드로 잰 값
라이선스
- 라이선스: Apache-2.0 — 원본 모델의 라이선스를 그대로 따름
- 권리: 원본 가중치·토크나이저·chat template은 Qwen 팀(Alibaba) 소유, 이 리포는 양자화 파생물