Skip to content

Switch Qwen3.5 FP4 SGLang configs to NVFP4 V2 / 将 Qwen3.5 FP4 SGLang 配置切换到 NVFP4 V2 #2205

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
adibarra merged 9 commits into from
Aug 10, 2026
Merged

Switch Qwen3.5 FP4 SGLang configs to NVFP4 V2 / 将 Qwen3.5 FP4 SGLang 配置切换到 NVFP4 V2 #2205

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
00ecba3
Switch Qwen3.5 FP4 configs to nvidia/Qwen3.5-397B-A17B-NVFP4-V2
hshrivastava-droid Jul 14, 2026
15cd958
Update perf-changelog pr-link for #2205
hshrivastava-droid Jul 14, 2026
25ef4a1
remove trt configs
hshrivastava-droid Jul 14, 2026
cac471c
Merge branch 'main' into qwen3.5-fp4-nvfp4-v2
hshrivastava-droid Jul 14, 2026
f85d3de
fix(runners): scope b200-dgxc qwen3.5 fp4 model path to the V2 checkp…
Ankur-singh Jul 29, 2026
f4b23a4
Merge origin/main and resolve perf changelog / 合并 origin/main 并解决 per…
Ankur-singh Jul 29, 2026
6562f5c
chore: merge origin/main and resolve perf changelog
Ankur-singh Jul 30, 2026
99e3a3f
Merge origin/main and resolve perf changelog / 合并 origin/main 并解决 per…
Ankur-singh Jul 30, 2026
b9e2555
chore: merge main into nvfp4 branch
adibarra Aug 10, 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
8 changes: 4 additions & 4 deletions configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1251,7 +1251,7 @@ qwen3.5-fp8-b200-sglang-agentic:

qwen3.5-fp4-b200-sglang:
image: lmsysorg/sglang:v0.5.14-cu130
model: nvidia/Qwen3.5-397B-A17B-NVFP4
model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
model-prefix: qwen3.5
runner: b200
precision: fp4
Expand All @@ -1267,7 +1267,7 @@ qwen3.5-fp4-b200-sglang:

qwen3.5-fp4-b200-sglang-mtp:
image: lmsysorg/sglang:v0.5.14-cu130
model: nvidia/Qwen3.5-397B-A17B-NVFP4
model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
model-prefix: qwen3.5
runner: b200
precision: fp4
Expand Down Expand Up @@ -1329,7 +1329,7 @@ qwen3.5-fp8-b300-sglang:

qwen3.5-fp4-b300-sglang:
image: lmsysorg/sglang:v0.5.14-cu130
model: nvidia/Qwen3.5-397B-A17B-NVFP4
model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
model-prefix: qwen3.5
runner: b300
precision: fp4
Expand All @@ -1345,7 +1345,7 @@ qwen3.5-fp4-b300-sglang:

qwen3.5-fp4-b300-sglang-mtp:
image: lmsysorg/sglang:v0.5.14-cu130
model: nvidia/Qwen3.5-397B-A17B-NVFP4
model: nvidia/Qwen3.5-397B-A17B-NVFP4-V2
model-prefix: qwen3.5
runner: b300
precision: fp4
Expand Down
10 changes: 10 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5281,3 +5281,13 @@
- "Enable multiple frontends and MessagePack request-plane encoding, and reduce 1P4D decode GPU memory utilization to 0.85"
- "Drop the recipe setup_script entirely (removes the DeepGEMM NVLink barrier timeout patch); required container deps now ship in vLLM 0.25.1, on NVIDIA/srt-slurm v1.0.31"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2139

- config-keys:
- qwen3.5-fp4-b200-sglang
- qwen3.5-fp4-b200-sglang-mtp
- qwen3.5-fp4-b300-sglang
- qwen3.5-fp4-b300-sglang-mtp
description:
- "Switch Qwen3.5 FP4 SGLang configs to the nvidia/Qwen3.5-397B-A17B-NVFP4-V2 checkpoint"
- "Model pre-staged on all clusters as Qwen3.5-397B-A17B-NVFP4-V2"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2205
10 changes: 10 additions & 0 deletions runners/launch_b200-dgxc.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -39,6 +39,16 @@ elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "bf16" ]]; then
elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp8" ]]; then
export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-FP8"
export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp8"
# qwen3.5 fp4 spans two checkpoints, so this must branch on the checkpoint and
# not on MODEL_PREFIX+PRECISION alone: the sglang keys moved to NVFP4-V2 while
# qwen3.5-fp4-b200-trt / -trt-mtp still declare plain NVFP4. Both share
# model-prefix qwen3.5 + precision fp4 + runner b200, and further down this
# script does `export MODEL="$MODEL_PATH"`, so a single shared branch would
# serve V2 weights to the TRT configs while publishing them under the old
# checkpoint name.
elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" && $MODEL == *NVFP4-V2 ]]; then
export MODEL_PATH="/scratch/fsw/models/Qwen3.5-397B-A17B-NVFP4-V2"
export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp4"
elif [[ $MODEL_PREFIX == "qwen3.5" && $PRECISION == "fp4" ]]; then
export MODEL_PATH="/lustre/fsw/models/Qwen3.5-397B-A17B-NVFP4"
export SRT_SLURM_MODEL_PREFIX="qwen3.5-fp4"
Expand Down
1 change: 1 addition & 0 deletions runners/launch_b300-nv.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -400,6 +400,7 @@ else
Qwen3.5-397B-A17B
Qwen3.5-397B-A17B-FP8
Qwen3.5-397B-A17B-NVFP4
Qwen3.5-397B-A17B-NVFP4-V2
gpt-oss-120b
)

Expand Down
Loading