Skip to content

Tune MiniMax-M3 AgentX on B200 and B300 with updated vLLM / 更新 B200 与 B300 上的 MiniMax-M3 AgentX 配置 #3435

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
xinli-sw wants to merge 12 commits into
base: main
Choose a base branch
from
+228 −38
Open

Tune MiniMax-M3 AgentX on B200 and B300 with updated vLLM / 更新 B200 与 B300 上的 MiniMax-M3 AgentX 配置 #3435

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
89d5ec0
perf: tune MiniMax-M3 AgentX on B200 and B300
xinli-sw Sep 25, 2026
78beae3
chore: link MiniMax-M3 AgentX changelog to PR
xinli-sw Sep 25, 2026
75630de
Merge branch 'main' into codex/minimaxm3-b200-b300-vllm-agentx
xinli-sw Sep 25, 2026
3b7cd43
test: handle git config flag in SRT clone stub
xinli-sw Sep 25, 2026
37d6595
fix: avoid FA4 FP8 draft KV startup failure
xinli-sw Sep 25, 2026
9a1fc30
fix: use FlashInfer for FP8 MiniMax EAGLE3 draft
xinli-sw Sep 25, 2026
f2f1ed8
Merge main into MiniMax AgentX update
xinli-sw Sep 25, 2026
849d364
fix: use FP8 Triton attention for MiniMax draft
xinli-sw Sep 26, 2026
baffa83
Merge main into MiniMax FP8 draft update
xinli-sw Sep 26, 2026
33fe121
fix: limit sweep changelog to MiniMax entries
xinli-sw Sep 26, 2026
dd8abfb
merge: sync main and port MiniMax AgentX recipes
xinli-sw Sep 26, 2026
baa8b19
fix: reserve B300 TP2 MSA workspace
xinli-sw Sep 26, 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
57 changes: 44 additions & 13 deletions benchmarks/single_node/srt-slurm-recipes/minimaxm3/vllm/b200-fp4-mtp/agentic.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -5,7 +5,7 @@ base:
name: minimaxm3-fp4-b200-vllm-agentic
model:
path: hf:nvidia/MiniMax-M3-NVFP4
container: vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45
container: vllm/vllm-openai:nightly-29468dde8b515031dc6d4d9d06bf0a2fa0442098
precision: fp4
resources:
gpu_type: b200
Expand All @@ -30,10 +30,11 @@ base:
workers: 1
args:
served-model-name: nvidia/MiniMax-M3-NVFP4
gpu-memory-utilization: 0.9
gpu-memory-utilization: 0.95
block-size: 128
language-model-only: true
enable-prefix-caching: true
enable-chunked-prefill: true
no-enable-flashinfer-autotune: true
reasoning-parser: minimax_m3
tool-call-parser: minimax_m3
Expand All @@ -47,7 +48,7 @@ base:
trust-remote-code: true
# Three-token EAGLE3 with the GQA draft head. Throughput runs switch to
# synthetic rejection at the golden acceptance length.
speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASH_ATTN"}'
speculative-config: '{"method":"eagle3","model":"Inferact/MiniMax-M3-EAGLE3-GQA","num_speculative_tokens":3,"attention_backend":"FLASHINFER"}'
env:
PYTHONNOUSERSITE: '1'
VLLM_ENGINE_READY_TIMEOUT_S: '3600'
Expand All @@ -61,13 +62,15 @@ base:
AIPERF_REQUIRED_SERVER_METRIC_PREFIX: 'vllm:'

# One variant per point. DRAM points give SimpleCPUOffload the whole host budget
# (TOTAL_CPU_DRAM_GB GiB) with lazy offload.
# (TOTAL_CPU_DRAM_GB decimal GB) with lazy offload.
override_tp8_c1:
roles:
agg:
gpus: 8
args:
tensor-parallel-size: 8
max-num-seqs: 2
max-cudagraph-capture-size: 8
benchmark:
env:
CONC: '1'
Expand All @@ -79,6 +82,8 @@ override_tp4_c1:
gpus: 4
args:
tensor-parallel-size: 4
max-num-seqs: 2
max-cudagraph-capture-size: 8
benchmark:
env:
CONC: '1'
Expand All @@ -90,6 +95,8 @@ override_tp4_c5:
gpus: 4
args:
tensor-parallel-size: 4
max-num-seqs: 10
max-cudagraph-capture-size: 40
benchmark:
env:
CONC: '5'
Expand All @@ -101,6 +108,8 @@ override_tp4_c10:
gpus: 4
args:
tensor-parallel-size: 4
max-num-seqs: 20
max-cudagraph-capture-size: 80
benchmark:
env:
CONC: '10'
Expand All @@ -112,6 +121,8 @@ override_tp4_c15:
gpus: 4
args:
tensor-parallel-size: 4
max-num-seqs: 30
max-cudagraph-capture-size: 120
benchmark:
env:
CONC: '15'
Expand All @@ -123,6 +134,8 @@ override_tp4_c20:
gpus: 4
args:
tensor-parallel-size: 4
max-num-seqs: 40
max-cudagraph-capture-size: 160
benchmark:
env:
CONC: '20'
Expand All @@ -134,7 +147,9 @@ override_tp4_c15_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 30
max-cudagraph-capture-size: 120
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -149,7 +164,9 @@ override_tp4_c20_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 40
max-cudagraph-capture-size: 160
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -164,7 +181,9 @@ override_tp4_c25_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 50
max-cudagraph-capture-size: 200
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -179,7 +198,9 @@ override_tp4_c30_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 60
max-cudagraph-capture-size: 240
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -194,7 +215,9 @@ override_tp4_c32_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 64
max-cudagraph-capture-size: 256
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -209,7 +232,9 @@ override_tp4_c34_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 68
max-cudagraph-capture-size: 272
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -224,7 +249,9 @@ override_tp4_c36_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 72
max-cudagraph-capture-size: 288
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -239,7 +266,9 @@ override_tp4_c38_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 76
max-cudagraph-capture-size: 304
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand All @@ -254,7 +283,9 @@ override_tp4_c40_dram:
gpus: 4
args:
tensor-parallel-size: 4
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1161788653568,"lazy_offload":true}}'
max-num-seqs: 80
max-cudagraph-capture-size: 320
kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":1082000000000,"lazy_offload":true}}'
env:
VLLM_USE_SIMPLE_KV_OFFLOAD: '1'
benchmark:
Expand Down
Loading
Loading