Skip to content

Add Qwen3.5 FP8 and NVFP4 #168

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
JustinTong0323 merged 5 commits into from
Mar 2, 2026
Merged

Add Qwen3.5 FP8 and NVFP4 #168

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
81cd5e0
Add Qwen3.5-397B-A17B-FP8
mmangkad Feb 18, 2026
4fdb29d
upd
mmangkad Feb 24, 2026
f1c2cd4
add nvfp4
mmangkad Feb 24, 2026
ef661ed
upd
mmangkad Feb 28, 2026
593a0f8
Update Qwen3.5 documentation to mark it as complete and reorganize ha…
JustinTong0323 Mar 2, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
282 changes: 282 additions & 0 deletions data/models/generated/v0.5.8/qwen35.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,3 +132,285 @@ families:
- '4'
prefill: null
decode: null
- name: Qwen3.5-397B-A17B-FP8
model_path: Qwen/Qwen3.5-397B-A17B-FP8
attributes:
llm:
thinking_capability: hybrid
tool_parser: qwen3_coder
reasoning_parser: qwen3
chat_template: null
hardware:
H100:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 8
dp: null
ep: null
enable_dp_attention: null
extra_args: []
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 8
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
H200:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args: []
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
B200:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args: []
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
B300:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 2
dp: null
ep: null
enable_dp_attention: null
extra_args: []
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp8
quantized_model_path: null
engine:
env_vars: {}
tp: 2
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
- name: Qwen3.5-397B-A17B-NVFP4
model_path: nvidia/Qwen3.5-397B-A17B-NVFP4
attributes:
llm:
thinking_capability: hybrid
tool_parser: qwen3_coder
reasoning_parser: qwen3
chat_template: null
hardware:
B200:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp4
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp4
quantized_model_path: null
engine:
env_vars: {}
tp: 4
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
B300:
configurations:
- name: default
attributes:
nodes: single
optimization: balanced
quantization: fp4
quantized_model_path: null
engine:
env_vars: {}
tp: 2
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
prefill: null
decode: null
- name: speculative-mtp
attributes:
nodes: single
optimization: low-latency
quantization: fp4
quantized_model_path: null
engine:
env_vars: {}
tp: 2
dp: null
ep: null
enable_dp_attention: null
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
- --speculative-algo
- NEXTN
- --speculative-num-steps
- '3'
- --speculative-eagle-topk
- '1'
- --speculative-num-draft-tokens
- '4'
prefill: null
decode: null
41 changes: 40 additions & 1 deletion data/models/src/v0.5.8/qwen35.yaml
View file
Open in desktop
Comment thread
mmangkad marked this conversation as resolved.
Original file line number Diff line number Diff line change
@@ -1,10 +1,18 @@
# Qwen3.5 Model Configurations (Simplified Format)
# This file is compiled to data/models/generated/qwen35.yaml
#
# GPU requirements:
# GPU requirements (BF16):
# H100: tp=16 (model ~800GB in BF16, each rank needs ~100GB > 80GB)
# H200: tp=8
# B200: tp=8
#
# GPU requirements (FP8):
# H100: tp=8 (model ~400GB in FP8, each rank needs ~50GB < 80GB)
# H200: tp=8
# B200: tp=8
#
# GPU requirements (FP4):
# B200 (183GB) and B300 (275GB) - FP4 requires Blackwell architecture

vendor: qwen

Expand Down Expand Up @@ -41,3 +49,34 @@ families:
models:
- name: Qwen3.5-397B-A17B
quantization: bf16

- name: Qwen3.5-397B-A17B-FP8
quantization: fp8
hardware:
H100: { tp: 8 }
H200: { tp: 4 }
B200: { tp: 4 }
B300: { tp: 2 }

- name: Qwen3.5-397B-A17B-NVFP4
quantization: fp4
model_path: nvidia/Qwen3.5-397B-A17B-NVFP4
hardware:
B200:
tp: 4
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
Comment on lines +72 to +73
Copy link
Copy Markdown

@hlu1 hlu1 Mar 2, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not just use the default value? flashinfer_cutlass is the current best due to an issue in flashinfer, but when the flashinfer issue is fixed, this will not be the best option.

B300:
tp: 2
extra_args:
- --attention-backend
- trtllm_mha
- --moe-runner-backend
- flashinfer_trtllm
- --fp4-gemm-backend
- flashinfer_cutlass
34 changes: 24 additions & 10 deletions docs/autoregressive/Qwen/Qwen3.5.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -18,6 +18,8 @@ Qwen3.5 features a Gated Delta Networks combined with sparse Mixture-of-Experts
**Available Models:**

- **BF16 (Full precision)**: [Qwen/Qwen3.5-397B-A17B](https://huggingface.co/Qwen/Qwen3.5-397B-A17B)
- **FP8 (8-bit Quantized)**: [Qwen/Qwen3.5-397B-A17B-FP8](https://huggingface.co/Qwen/Qwen3.5-397B-A17B-FP8)
- **FP4 (4-bit Quantized)**: [nvidia/Qwen3.5-397B-A17B-NVFP4](https://huggingface.co/nvidia/Qwen3.5-397B-A17B-NVFP4)

**License:** Apache 2.0

Expand Down Expand Up @@ -49,22 +51,34 @@ import Qwen35ConfigGenerator from '@site/src/components/autoregressive/Qwen35Con

### 3.2 Configuration Tips

- The model has ~397B parameters in BF16, requiring ~800GB of GPU memory for weights alone.
- **H100 (80GB)** requires tp=16 (2 nodes) since each rank needs ~100GB at tp=8.
- **H200 (141GB)** and **B200 (192GB)** can run with tp=8 on a single node.
- Speculative decoding (MTP) can significantly reduce latency for interactive use cases.
- The `--mem-fraction-static` flag is recommended for optimal memory utilization, adjust it based on your hardware and workload.
- Context length defaults to 262,144 tokens. If you encounter OOM errors, consider reducing it, but maintain at least 128K to preserve thinking capabilities.
- To speed up weight loading for this large model, add `--model-loader-extra-config='{"enable_multithread_load": "true","num_threads": 64}'` to the launch command.
- **CUDA IPC Transport**: Add `SGLANG_USE_CUDA_IPC_TRANSPORT=1` as an environment variable to use CUDA IPC for transferring multimodal features, significantly improving TTFT (Time To First Token). Note: this consumes additional memory proportional to image size, so you may need to lower `--mem-fraction-static` or `--max-running-requests`.
- **Multimodal Attention Backend**: Use `--mm-attention-backend fa3` on H100/H200 for better vision performance, or `--mm-attention-backend fa4` on B200.
- **Multimodal Attention Backend**: Use `--mm-attention-backend fa3` on H100/H200 for better vision performance, or `--mm-attention-backend fa4` on B200/B300.
- For processing large images or videos, you may need to lower `--mem-fraction-static` to leave room for image feature tensors.

| Hardware | TP |
| -------- | -- |
| H100 | 16 |
| H200 | 8 |
| B200 | 8 |
- Hardware requirements:
- **BF16**: ~397B parameters require ~800GB of GPU memory for weights.
- **H100 (80GB)** requires tp=16 (2 nodes) since each rank needs ~100GB at tp=8.
- **H200 (141GB)** runs with tp=8.
- **B200 (183GB)** runs with tp=8.
- **B300 (275GB)** runs with tp=4.
- **FP8**: The FP8 quantized model requires ~400GB for weights, cutting memory in half.
- **H100 (80GB)** runs with tp=8.
- **H200 (141GB)** runs with tp=4.
- **B200 (183GB)** runs with tp=4.
- **B300 (275GB)** runs with tp=2.
- **FP4**: The FP4 quantized model requires ~250GB for weights, cutting memory by almost 4x. Only compatible with B200/B300 (Blackwell architecture).
- **B200 (183GB)** runs with tp=4.
- **B300 (275GB)** runs with tp=2.

| Hardware | Memory | BF16 TP | FP8 TP | FP4 TP |
| -------- | ------ | ------- | ------ | --------------- |
| H100 | 80GB | 16 | 8 | N/A |
| H200 | 141GB | 8 | 4 | N/A |
| B200 | 183GB | 8 | 4 | 4 |
| B300 | 275GB | 4 | 2 | 2 |

## 4. Model Invocation

Expand Down
Loading