Skip to content

[Do Not Merge] kimik2.5-fp4-b300-vllm: align server launch with B200 recipe #1698

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
RohitNagraj wants to merge 4 commits into
base: main
Choose a base branch
from
Open

[Do Not Merge] kimik2.5-fp4-b300-vllm: align server launch with B200 recipe #1698

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
a1eb2ed
perf: align kimik2.5-fp4-b300-vllm server launch with B200 recipe
RohitNagraj Jun 9, 2026
62c377e
chore: set perf-changelog pr-link for kimik2.5-fp4-b300-vllm server t…
RohitNagraj Jun 9, 2026
3274c53
Changed max-cudagraph-capture-size
RohitNagraj Jun 10, 2026
993085b
perf: tune kimik2.5-fp4-b300-vllm server launch (batched-tokens=cudag…
RohitNagraj Jun 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
11 changes: 10 additions & 1 deletion benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,6 +38,12 @@ nvidia-smi
export TORCH_CUDA_ARCH_LIST="10.0"
export PYTHONNOUSERSITE=1

# vLLM v0.20.2+'s CUDA-graph memory profiler pre-reserves a large chunk of GPU
# memory upfront, which collides with --gpu-memory-utilization=0.90 and shrinks
# the effective budget left for the KV cache. Disable the profiler so 0.90 means
# 0.90 (same pattern as benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b200.sh).
export VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0

SERVER_LOG=/workspace/server.log

if [ "${EVAL_ONLY}" = "true" ]; then
Expand All @@ -56,7 +62,10 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
--reasoning-parser kimi_k2 \
--tool-call-parser kimi_k2 \
--compilation_config.pass_config.fuse_allreduce_rms true \
--no-enable-prefix-caching \
--kv-cache-dtype fp8 \
--max-cudagraph-capture-size 8192 \
Comment thread
cursor[bot] marked this conversation as resolved.
--max-num-batched-tokens 8192 \
--stream-interval 20 --no-enable-prefix-caching \
--trust-remote-code > $SERVER_LOG 2>&1 &

SERVER_PID=$!
Expand Down
6 changes: 6 additions & 0 deletions perf-changelog.yaml
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3531,3 +3531,9 @@
- "The Rust frontend replaces only the Python serving/API layer (HTTP, tokenization, scheduling glue, detokenization) and spawns the same Python EngineCore, so GPU kernels/attention/MoE GEMM/KV cache are untouched"
- "A/B sweep (28 single-node points, 1k1k + 8k1k, TP 1/2/4) vs the Python-frontend baseline (run 26696260751): throughput Pareto-neutral (peak tok/s/GPU within <1.5%, frontiers coincident) and TPOT flat (+-0.5%); TTFT improves ~8% at 1k1k and ~22% at 8k1k (every point), the expected signature of lower frontend CPU latency before first token, scaling with input length"
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1634

- config-keys:
- kimik2.5-fp4-b300-vllm
description:
- "Tune the B300 vLLM server launch: set --kv-cache-dtype fp8, --max-cudagraph-capture-size 8192, --max-num-batched-tokens 8192 (matched to the cudagraph capture size), and --stream-interval 20; disable the CUDA-graph memory profiler (VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS=0) so --gpu-memory-utilization is honored in full."
pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1698