diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml index 0a6c86df6d..298c8babfd 100644 --- a/configs/nvidia-master.yaml +++ b/configs/nvidia-master.yaml @@ -7217,10 +7217,10 @@ qwen3.5-fp8-b300-sglang-agentic-mtp: - { tp: 2, ep: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: hicache }, conc-list: [32, 34, 36, 38, 40, 44, 48, 52, 56] } qwen3.5-fp4-b300-sglang-agentic-mtp: - image: lmsysorg/sglang:v0.5.16-cu130 + image: lmsysorg/sglang:nightly-dev-cu13-20260907-30705c00 model: nvidia/Qwen3.5-397B-A17B-NVFP4 model-prefix: qwen3.5 - runner: cluster:b300-nv + runner: cluster:b300-dsxe precision: fp4 framework: sglang multinode: false diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 6b249ea15e..5e599e82aa 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -6921,3 +6921,12 @@ - "Pick up the latest automatic ROCm DeepSeek-V4 optimizations, including fused mHC post/pre plus RMSNorm, gfx950 C4A top-k dispatch, fused C4 compressor GEMMs, fused SWA q/kv RMSNorm plus q FP8 quantization, and medium-batch cooperative top-k tuning." - "Keep the existing VLLM_ROCM_USE_AITER=1, VLLM_ROCM_USE_AITER_MOE=1, and --moe-backend aiter settings, and explicitly add VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 plus VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 to both STP and MTP paths. The current checkpoint's shared-expert path does not satisfy the latest vLLM fusion conditions, so that fusion flag self-disables while preserving recipe parity." pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2792 + +- config-keys: + - qwen3.5-fp4-b300-sglang-agentic-mtp + scenario-type: + - agentic-coding + description: + - "Update SGLang image from lmsysorg/sglang:v0.5.16-cu130 (v0.5.16 release) to lmsysorg/sglang:nightly-dev-cu13-20260907-30705c00 (2026-09-07 cu13 dev nightly, digest sha256:19b8fa1223cc339c1eae7a5b703f1a8c2543b5b119155bf3d7efaef18f77f007, tag commit sgl-project/sglang@30705c00; Docker Hub last pushed 2026-09-07T01:43:42Z), the same tag the Qwen3.5 SGLang AgentX recipes on B200/H200/H100 moved to in #2861/#2862/#2868/#2869. benchmarks/single_node/agentic/qwen3.5_fp4_b300_sglang_mtp.sh is unchanged: modelopt_fp4 quantization, fp8_e4m3 KV, trtllm_mha attention, flashinfer_trtllm MoE runner, native NEXTN MTP with golden AL 3.39; the TP4 and TP2/EP2 grids are unchanged. The qwen3.5-fp4-b300-sglang-agentic-power-ab key is deliberately not bumped: it is a controlled FP8/FP4 power matrix that requires an identical SGLang build across both precisions." + - "Move the recipe from the retired cluster:b300-nv fleet (launcher and runner labels removed in #2826) to cluster:b300-dsxe so the sweep has a runner to schedule on. The runner change means this is not an append-only bump: the whole curve reruns on DSXE." + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2882