From 09a23b0b44d26fd1930108f61f9b1b4994f75be0 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Sun, 17 May 2026 19:41:54 -0400 Subject: [PATCH 1/5] Update glm5-fp8-b200-sglang (+mtp) SGLang image to v0.5.12-cu130 Drop the 33/29-day-old pinned nightly lmsysorg/sglang:nightly-dev-cu13-20260317-1eea7448 in favor of the stable v0.5.12-cu130 release tag used by other b200 cu130 sglang recipes on main. --- .github/configs/nvidia-master.yaml | 6 +++--- perf-changelog.yaml | 7 +++++++ 2 files changed, 10 insertions(+), 3 deletions(-) diff --git a/.github/configs/nvidia-master.yaml b/.github/configs/nvidia-master.yaml index 37dd5af3f2..4c1fde805f 100644 --- a/.github/configs/nvidia-master.yaml +++ b/.github/configs/nvidia-master.yaml @@ -2188,7 +2188,7 @@ qwen3.5-fp4-b200-sglang-mtp: - { tp: 2, ep: 1, conc-start: 4, conc-end: 64, spec-decoding: mtp } glm5-fp8-b200-sglang: - image: lmsysorg/sglang:nightly-dev-cu13-20260317-1eea7448 + image: lmsysorg/sglang:v0.5.12-cu130 model: zai-org/GLM-5-FP8 model-prefix: glm5 runner: b200 @@ -2207,7 +2207,7 @@ glm5-fp8-b200-sglang: - { tp: 8, ep: 1, conc-start: 4, conc-end: 256 } glm5-fp8-b200-sglang-mtp: - image: lmsysorg/sglang:nightly-dev-cu13-20260317-1eea7448 + image: lmsysorg/sglang:v0.5.12-cu130 model: zai-org/GLM-5-FP8 model-prefix: glm5 runner: b200 @@ -2234,7 +2234,7 @@ glm5-fp8-b200-sglang-mtp: # (either main had none or had a different conc/offload sweep). # The original glm5-fp8-b200-sglang entry stays byte-identical to origin/main. glm5-fp8-b200-sglang-agentic: - image: lmsysorg/sglang:nightly-dev-cu13-20260317-1eea7448 + image: lmsysorg/sglang:v0.5.12-cu130 model: zai-org/GLM-5-FP8 model-prefix: glm5 runner: b200 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 13a695453a..8184ab13c4 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -2629,3 +2629,10 @@ description: - "Update vLLM ROCm image from v0.18.0 to v0.21.0" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1404 + +- config-keys: + - glm5-fp8-b200-sglang + - glm5-fp8-b200-sglang-mtp + description: + - "Update SGLang image from nightly-dev-cu13-20260317-1eea7448 (33d/29d old) to v0.5.12-cu130" + pr-link: PLACEHOLDER From 6c3182d4950372bbaacc05a4f59251d962654001 Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Sun, 17 May 2026 19:41:57 -0400 Subject: [PATCH 2/5] chore: fill pr-link for #1447 --- perf-changelog.yaml | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 8184ab13c4..6c56a160ab 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -2635,4 +2635,4 @@ - glm5-fp8-b200-sglang-mtp description: - "Update SGLang image from nightly-dev-cu13-20260317-1eea7448 (33d/29d old) to v0.5.12-cu130" - pr-link: PLACEHOLDER + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1447 From ffce090813141cb7f65db854c1815fe58474a7e9 Mon Sep 17 00:00:00 2001 From: claude-fix-bot Date: Sun, 17 May 2026 21:50:03 -0400 Subject: [PATCH 3/5] Re-trigger sweep (abandon in-flight run with transient infra failures) From 739585b067f2e83e6a3190019a002d64e0bae431 Mon Sep 17 00:00:00 2001 From: claude-fix-bot Date: Mon, 18 May 2026 03:30:34 -0400 Subject: [PATCH 4/5] fix(glm5_fp8_b200): use --fp8-gemm-runner-backend cutlass MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Bypass the DeepGemm CUDA_ERROR_ILLEGAL_ADDRESS (700) crash at runtime_utils.hpp:143 during CUDA-graph capture. Previously documented on B300 (KLAUD_DEBUG §4a) and now confirmed on B200 (sm_100) too at high concurrency — spec-none conc=128 hit the same illegal-memory- access during graph capture. cutlass is the documented fallback backend and works on both sm_100 and sm_103. --- benchmarks/single_node/glm5_fp8_b200.sh | 1 + benchmarks/single_node/glm5_fp8_b200_mtp.sh | 1 + 2 files changed, 2 insertions(+) diff --git a/benchmarks/single_node/glm5_fp8_b200.sh b/benchmarks/single_node/glm5_fp8_b200.sh index 2cd84dddc7..ad47210d25 100755 --- a/benchmarks/single_node/glm5_fp8_b200.sh +++ b/benchmarks/single_node/glm5_fp8_b200.sh @@ -45,6 +45,7 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0. --tool-call-parser glm47 \ --reasoning-parser glm45 \ --kv-cache-dtype fp8_e4m3 --quantization fp8 \ +--fp8-gemm-runner-backend cutlass \ --attention-backend nsa \ --nsa-decode-backend trtllm --nsa-prefill-backend trtllm \ --moe-runner-backend flashinfer_trtllm \ diff --git a/benchmarks/single_node/glm5_fp8_b200_mtp.sh b/benchmarks/single_node/glm5_fp8_b200_mtp.sh index ecd5ca0af7..6262e944cf 100755 --- a/benchmarks/single_node/glm5_fp8_b200_mtp.sh +++ b/benchmarks/single_node/glm5_fp8_b200_mtp.sh @@ -46,6 +46,7 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0. --tool-call-parser glm47 \ --reasoning-parser glm45 \ --kv-cache-dtype fp8_e4m3 --quantization fp8 \ +--fp8-gemm-runner-backend cutlass \ --attention-backend nsa \ --nsa-decode-backend trtllm --nsa-prefill-backend trtllm \ --moe-runner-backend flashinfer_trtllm \ From d40fd278b8ab627122ac0e88f93dd5f9f72a5f15 Mon Sep 17 00:00:00 2001 From: claude-fix-bot Date: Mon, 18 May 2026 12:55:08 -0400 Subject: [PATCH 5/5] fix(glm5_fp8_b200): correct flag name --fp8-gemm-backend (not --fp8-gemm-runner-backend) MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Previous attempt used --fp8-gemm-runner-backend which sglang v0.5.12 rejects with 'sglang serve: error: unrecognized arguments'. The actual flag is --fp8-gemm-backend (confirmed via the help dump in the failing job's log). KLAUD_DEBUG.md §4a had the wrong name; needs a separate fix-up PR. --- benchmarks/single_node/glm5_fp8_b200.sh | 2 +- benchmarks/single_node/glm5_fp8_b200_mtp.sh | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/benchmarks/single_node/glm5_fp8_b200.sh b/benchmarks/single_node/glm5_fp8_b200.sh index ad47210d25..ccaa87b980 100755 --- a/benchmarks/single_node/glm5_fp8_b200.sh +++ b/benchmarks/single_node/glm5_fp8_b200.sh @@ -45,7 +45,7 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0. --tool-call-parser glm47 \ --reasoning-parser glm45 \ --kv-cache-dtype fp8_e4m3 --quantization fp8 \ ---fp8-gemm-runner-backend cutlass \ +--fp8-gemm-backend cutlass \ --attention-backend nsa \ --nsa-decode-backend trtllm --nsa-prefill-backend trtllm \ --moe-runner-backend flashinfer_trtllm \ diff --git a/benchmarks/single_node/glm5_fp8_b200_mtp.sh b/benchmarks/single_node/glm5_fp8_b200_mtp.sh index 6262e944cf..5e4f98533d 100755 --- a/benchmarks/single_node/glm5_fp8_b200_mtp.sh +++ b/benchmarks/single_node/glm5_fp8_b200_mtp.sh @@ -46,7 +46,7 @@ PYTHONNOUSERSITE=1 python3 -m sglang.launch_server --model-path=$MODEL --host=0. --tool-call-parser glm47 \ --reasoning-parser glm45 \ --kv-cache-dtype fp8_e4m3 --quantization fp8 \ ---fp8-gemm-runner-backend cutlass \ +--fp8-gemm-backend cutlass \ --attention-backend nsa \ --nsa-decode-backend trtllm --nsa-prefill-backend trtllm \ --moe-runner-backend flashinfer_trtllm \