Skip to content

[Klaud Cold] Update dsv4-fp4-b300-vllm (+mtp) vLLM image to v0.22.0 #1595

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

Open
functionstackx wants to merge 1 commit into
base: main
Choose a base branch
from
+9 −2
Open

[Klaud Cold] Update dsv4-fp4-b300-vllm (+mtp) vLLM image to v0.22.0 #1595

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
4 changes: 2 additions & 2 deletions .github/configs/nvidia-master.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3003,7 +3003,7 @@ dsv4-fp8-h200-sglang-mtp:
# field, so dp-attn=true is used as the existing vLLM script switch for DP4
# layouts on 4 allocated GPUs.
dsv4-fp4-b300-vllm:
image: vllm/vllm-openai:v0.21.0
image: vllm/vllm-openai:v0.22.0
model: deepseek-ai/DeepSeek-V4-Pro
model-prefix: dsv4
runner: b300
Expand Down Expand Up @@ -3099,7 +3099,7 @@ dsv4-fp4-b300-trt-mtp:
- { tp: 8, ep: 8, dp-attn: true, conc-start: 256, conc-end: 1024, spec-decoding: mtp }

dsv4-fp4-b300-vllm-mtp:
image: vllm/vllm-openai:v0.21.0
image: vllm/vllm-openai:v0.22.0
model: deepseek-ai/DeepSeek-V4-Pro
model-prefix: dsv4
runner: b300
Expand Down
7 changes: 7 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3239,3 +3239,10 @@
- "Add --compilation-config '{\"mode\":3,\"cudagraph_mode\":\"PIECEWISE\"}' to vllm serve, mirroring `model.base_args` from the upstream recipe. `pass_config.fuse_minimax_qk_norm` from the recipe is intentionally omitted — it triggers an upstream NameError on ROCm because vllm/compilation/passes/pass_manager.py imports MiniMaxQKNormPass under `is_cuda()` (NVIDIA-only) while using it unconditionally"
- "Conditionally enable VLLM_ROCM_SHUFFLE_KV_CACHE_LAYOUT=1 per (TP, EP, CONC) — on for shapes where the AITER ASM paged-attention kernel exists in the gfx942 heuristic table (TP=2 EP=1 CONC<=16, TP=8 EP=8 CONC<=64), off otherwise. Above the thresholds vllm/v1/attention/backends/rocm_aiter_fa.py routes decode through aiter pa_fwd_asm and crashes with `RuntimeError: get_heuristic_kernel: cannot get heuristic kernel!` for MiniMax-M2.5's attention shape (gqa=6 block_size=32 qTile=0); below them the ASM auto-dispatch is the perf win the recipe wants. Thresholds confirmed across 17 bench cells + 3 eval cells in PR #1594 sweep run 26692603804. Mirrors the per-shape toggle pattern in benchmarks/single_node/minimaxm2.5_fp8_mi355x.sh; can collapse to unconditional SHUFFLE=1 once AITER registers the missing kernel on gfx942"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1594

- config-keys:
- dsv4-fp4-b300-vllm
- dsv4-fp4-b300-vllm-mtp
description:
- "Update vLLM image from v0.21.0 to v0.22.0"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1595
Loading