Skip to content

perf(dsv4-fp4-mi355x-vllm): use AITER a16w4 MoE backend (+21% decode) #1989

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

Draft
jiacao-amd wants to merge 1 commit into
base: main
Choose a base branch
from
Draft

perf(dsv4-fp4-mi355x-vllm): use AITER a16w4 MoE backend (+21% decode) #1989

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
16 changes: 10 additions & 6 deletions benchmarks/single_node/fixed_seq_len/dsv4_fp4_mi355x_vllm.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -12,11 +12,14 @@ set -eo pipefail
# same ROCm recipe while switching parallelism to vLLM's DP+EP form.
# Image-pin details live in amd-master.yaml.
#
# --moe-backend triton_unfused is required for the FP4 MoE expert
# weight format used by deepseek-ai/DeepSeek-V4-Pro. Letting --moe-backend
# default to auto picks a backend that doesn't register the FP4 scale
# parameters (w13_weight_scale / w2_weight_scale), so safetensors
# loading raises KeyError.
# --moe-backend aiter selects the AITER AITER_MXFP4_BF16 (a16w4) MoE
# kernel, which is ~21% faster on decode than triton_unfused
# (1105 vs 915 output tok/s at c64 1k/1k, TP=8) while staying numerically
# stable. It requires VLLM_ROCM_USE_AITER=1 and VLLM_ROCM_USE_AITER_MOE=1
# (both exported below).
# Do NOT let --moe-backend default to auto: auto picks a backend that
# doesn't register the FP4 scale parameters (w13_weight_scale /
# w2_weight_scale), so safetensors loading raises KeyError.
#
# --compilation-config mode=3 with FULL_AND_PIECEWISE cudagraph mode
# enables full CUDA graph capture for improved throughput on MI355X.
Expand Down Expand Up @@ -45,6 +48,7 @@ if [ -n "$ROCR_VISIBLE_DEVICES" ]; then
fi

export VLLM_ROCM_USE_AITER=1
export VLLM_ROCM_USE_AITER_MOE=1

SERVER_LOG=/workspace/server.log

Expand Down Expand Up @@ -75,7 +79,7 @@ vllm serve $MODEL --port $PORT \
--gpu-memory-utilization 0.8 \
--kv-cache-dtype fp8 \
--trust-remote-code \
--moe-backend triton_unfused \
--moe-backend aiter \
--tokenizer-mode deepseek_v4 \
--reasoning-parser deepseek_v4 \
--compilation-config '{"mode":3,"cudagraph_mode":"FULL_AND_PIECEWISE"}' > $SERVER_LOG 2>&1 &
Expand Down