Skip to content

[Klaud Cold] minimaxm3-fp8-mi355x-vllm-mtp: run with CUDA graphs (drop --enforce-eager) #1755

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
functionstackx merged 1 commit into from
Jun 14, 2026
Merged

[Klaud Cold] minimaxm3-fp8-mi355x-vllm-mtp: run with CUDA graphs (drop --enforce-eager) #1755

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
10 changes: 7 additions & 3 deletions benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_mi355x_mtp.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,8 +5,9 @@
# minimaxm3_fp8_mi355x.sh. Adds the Inferact/MiniMax-M3-EAGLE3 draft head via
# --speculative-config with 3 speculative tokens. Everything else mirrors the
# non-MTP recipe: MXFP8 from TP=4 on gfx950, mandatory --block-size 128,
# --language-model-only for the text-only benchmark, FP8 KV cache,
# --attention-backend TRITON_ATTN, and --enforce-eager.
# --language-model-only for the text-only benchmark, FP8 KV cache, and
# --attention-backend TRITON_ATTN. Runs with CUDA graphs (no --enforce-eager);
# VLLM_USE_BREAKABLE_CUDAGRAPH=0 avoids the M3-decode breakable-cudagraph path.
#
# Unlike the CUDA recipes, the drafter needs no attention_backend override:
# the FlashInfer "page size 128 requires GQA/MQA" limitation that forced
Expand Down Expand Up @@ -57,6 +58,9 @@ fi

SERVER_LOG=/workspace/server.log
export VLLM_ENGINE_READY_TIMEOUT_S=3600
# Run with CUDA graphs (no --enforce-eager): VLLM_USE_BREAKABLE_CUDAGRAPH=0
# avoids the M3-decode breakable-cudagraph path that previously forced eager.
export VLLM_USE_BREAKABLE_CUDAGRAPH=0

if [ "${EVAL_ONLY}" = "true" ]; then
setup_eval_context
Expand Down Expand Up @@ -168,11 +172,11 @@ set -x
vllm serve "$MODEL" --port "$PORT" \
"${PARALLEL_ARGS[@]}" \
--block-size 128 \
--no-enable-prefix-caching \
--language-model-only \
--max-model-len "$MAX_MODEL_LEN" \
--kv-cache-dtype fp8 \
--attention-backend TRITON_ATTN \
--enforce-eager \

@cursor cursor Bot Jun 14, 2026

Copy link
Copy Markdown

Choose a reason for hiding this comment

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

Missing MTP cudagraph capture sizing

Medium Severity

The EAGLE3 MTP recipe now runs with CUDA graphs but never passes --max-cudagraph-capture-size, while other MiniMax MTP scripts size capture to CONC * (1 + speculative tokens). Sweeps up to hundreds of concurrent requests can decode outside the captured graph sizes, so results may not reflect the intended full-graph speedup.

Fix in Cursor Fix in Web

Reviewed by Cursor Bugbot for commit 234abd2. Configure here.

--speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS}" \
--tool-call-parser minimax_m3 \
--reasoning-parser minimax_m3 \
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 @@ -3783,3 +3783,10 @@
- "H200-aligned layouts and concurrency ranges: TP4 and TP8 latency, TP4/TP8 expert parallelism, and TP8 data-parallel attention across 1k1k and 8k1k"
- "Route the MI325X Hugging Face cache and runtime compiler caches to node-local storage, and mount ROCm GPU devices explicitly"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1748

- config-keys:
- minimaxm3-fp8-mi355x-vllm-mtp
description:
- "Run the MiniMax-M3 MXFP8 MI355X EAGLE3 MTP recipe with CUDA graphs instead of --enforce-eager"
- "Drop --enforce-eager and set VLLM_USE_BREAKABLE_CUDAGRAPH=0, which avoids the M3-decode breakable-cudagraph path that previously forced eager execution (the non-MTP MI355X recipe already got this in #1754)"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1755