From 1d24294ceff704aee6a07be49672480c3c402147 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Tue, 14 Jul 2026 22:31:44 +0000 Subject: [PATCH 1/9] perf(amd): tune Kimi K2.5 MXFP4 on MI355X [skip-sweep] MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Apply the accuracy-gated serving settings, preserve the existing TP4 and TP8 matrix dimensions, extend only their concurrency upper bounds to 128, and keep the image as a draft placeholder pending the required AITER bump. 中文：应用已通过精度验证的服务配置，保留现有 TP4 和 TP8 矩阵维度，仅将并发上限扩展到 128，并暂时保留镜像占位符，等待所需的 AITER 版本升级。 --- .../fixed_seq_len/kimik2.5_fp4_mi355x.sh | 21 ++++++++++++++----- configs/amd-master.yaml | 11 +++++----- perf-changelog.yaml | 7 +++++++ 3 files changed, 29 insertions(+), 10 deletions(-) diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh index d4616143ac..c159af6a9f 100755 --- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh +++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh @@ -35,6 +35,10 @@ if [ "${EVAL_ONLY}" = "true" ]; then MAX_MODEL_LEN="$EVAL_MAX_MODEL_LEN" fi +if [ "$MAX_MODEL_LEN" -lt 9472 ]; then + MAX_MODEL_LEN=9472 +fi + # If the machine runs a MEC FW older than 177, RCCL # cannot reclaim some memory. # Disable that features to avoid crashes. @@ -47,6 +51,13 @@ fi export VLLM_ROCM_USE_AITER=1 export VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 +export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 +export VLLM_ROCM_USE_SKINNY_GEMM=0 +export AITER_MXFP4_INTERMEDIATE=1 +export AITER_BYPASS_TUNE_CONFIG=0 +export AITER_MOE_SORT_BACKEND=auto +export AITER_LOG_LEVEL=INFO +export OMP_NUM_THREADS=1 # Disable AITER RMSNorm for TP < 8 due to accuracy issues if [ "${TP}" -lt 8 ]; then @@ -59,9 +70,6 @@ else EP=" " fi -# following AMD andy luo's recipe -# https://x.com/linluo77/status/2017024513595301985 - # Start GPU monitoring (power, temperature, clocks every second) start_gpu_monitor @@ -71,8 +79,11 @@ vllm serve $MODEL --port $PORT \ $EP \ --gpu-memory-utilization 0.90 \ --max-model-len $MAX_MODEL_LEN \ ---block-size=1 \ ---no-enable-prefix-caching \ +--kv-cache-dtype fp8 \ +--block-size 16 \ +--max-num-batched-tokens 16384 \ +--max-num-seqs 512 \ +--async-scheduling \ --trust-remote-code \ --no-enable-prefix-caching \ --mm-encoder-tp-mode data > $SERVER_LOG 2>&1 & diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 4b1ea4aed3..2327426cce 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -770,7 +770,8 @@ kimik2.5-int4-mi300x-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-mi355x-vllm: - image: vllm/vllm-openai-rocm:v0.24.0 + # TODO: replace with the official vLLM ROCm image containing AITER #3832. + image: TBD model: amd/Kimi-K2.5-MXFP4 model-prefix: kimik2.5 runner: mi355x @@ -782,13 +783,13 @@ kimik2.5-fp4-mi355x-vllm: - isl: 1024 osl: 1024 search-space: - - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 4, conc-start: 4, conc-end: 64 } + - { tp: 8, conc-start: 4, conc-end: 128 } + - { tp: 4, conc-start: 4, conc-end: 128 } - isl: 8192 osl: 1024 search-space: - - { tp: 8, conc-start: 4, conc-end: 64 } - - { tp: 4, conc-start: 4, conc-end: 64 } + - { tp: 8, conc-start: 4, conc-end: 128 } + - { tp: 4, conc-start: 4, conc-end: 128 } kimik2.5-fp4-mi355x-vllm-agentic: image: vllm/vllm-openai-rocm:v0.22.0 diff --git a/perf-changelog.yaml b/perf-changelog.yaml index d77c35f1fd..c2151d778a 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -4781,3 +4781,10 @@ - "Bump image to lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260708" - "Clean the export envs" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2198 + +- config-keys: + - kimik2.5-fp4-mi355x-vllm + description: + - "Apply the accuracy-gated Kimi-K2.5 MXFP4 TP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, and the AITER BF16 GEMM path" + - "Extend the existing TP4 and TP8 concurrency sweeps from 64 to 128 without changing other search dimensions; the image remains TBD until an official vLLM ROCm image contains AITER #3832" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 From 601eefc91f79ec212f27d0dd50135b764ca58ba6 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Tue, 14 Jul 2026 18:55:19 -0400 Subject: [PATCH 2/9] Update kimik2.5_fp4_mi355x.sh --- benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh | 1 - 1 file changed, 1 deletion(-) diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh index c159af6a9f..e01ee4cfae 100755 --- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh +++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh @@ -55,7 +55,6 @@ export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 export VLLM_ROCM_USE_SKINNY_GEMM=0 export AITER_MXFP4_INTERMEDIATE=1 export AITER_BYPASS_TUNE_CONFIG=0 -export AITER_MOE_SORT_BACKEND=auto export AITER_LOG_LEVEL=INFO export OMP_NUM_THREADS=1 From 335b39a963f10da4b57dd917fb123a6eef1f79ce Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Tue, 14 Jul 2026 18:56:13 -0400 Subject: [PATCH 3/9] Update kimik2.5_fp4_mi355x.sh --- benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh index e01ee4cfae..6dadc26876 100755 --- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh +++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh @@ -55,7 +55,7 @@ export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 export VLLM_ROCM_USE_SKINNY_GEMM=0 export AITER_MXFP4_INTERMEDIATE=1 export AITER_BYPASS_TUNE_CONFIG=0 -export AITER_LOG_LEVEL=INFO +export AITER_MOE_SORT_BACKEND=auto export OMP_NUM_THREADS=1 # Disable AITER RMSNorm for TP < 8 due to accuracy issues From 7f57919036ec9f6ebbd175cd8af6c6ef7273467f Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Tue, 21 Jul 2026 17:45:04 +0000 Subject: [PATCH 4/9] perf(amd): pin rocm/vllm-dev:nightly_cdna4 for Kimi-K2.5 MXFP4 MI355X vLLM MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Replace image: TBD with rocm/vllm-dev:nightly_cdna4, the latest gfx950 serving image. It ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend required by this recipe. 中文：将 image: TBD 替换为最新的 gfx950 服务镜像 rocm/vllm-dev:nightly_cdna4。 该镜像内置 AITER 0.1.19.dev（不低于 vllm-project/vllm#48683 引入的 v0.1.16.post5）， 包含本配方所需的 ROCm/aiter#3832 gfx950 MXFP4 MoE 后端。 Co-authored-by: Cursor --- configs/amd-master.yaml | 5 +++-- 1 file changed, 3 insertions(+), 2 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 2327426cce..8692945c09 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -770,8 +770,9 @@ kimik2.5-int4-mi300x-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-mi355x-vllm: - # TODO: replace with the official vLLM ROCm image containing AITER #3832. - image: TBD + # rocm/vllm-dev:nightly_cdna4 ships AITER 0.1.19.dev (>= v0.1.16.post5 from + # vllm-project/vllm#48683) which includes the ROCm/aiter#3832 gfx950 MXFP4 MoE backend. + image: rocm/vllm-dev:nightly_cdna4 model: amd/Kimi-K2.5-MXFP4 model-prefix: kimik2.5 runner: mi355x From 2102c99c643f146d9f28f35bb6ea198030cdabf5 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Tue, 21 Jul 2026 17:58:12 +0000 Subject: [PATCH 5/9] perf(amd): pin immutable nightly_cdna4 build tag for Kimi-K2.5 MXFP4 MI355X vLLM MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Use the reproducible dated build tag rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 (same digest as floating nightly_cdna4) instead of the moving nightly_cdna4 tag. 中文：改用可复现的带日期构建标签 rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 （与浮动的 nightly_cdna4 同一 digest），替换会被覆盖的 nightly_cdna4 标签。 Co-authored-by: Cursor --- configs/amd-master.yaml | 4 +--- perf-changelog.yaml | 2 +- 2 files changed, 2 insertions(+), 4 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index 8733ae4320..f0eeb3b849 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -509,9 +509,7 @@ kimik2.5-int4-mi300x-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-mi355x-vllm: - # rocm/vllm-dev:nightly_cdna4 ships AITER 0.1.19.dev (>= v0.1.16.post5 from - # vllm-project/vllm#48683) which includes the ROCm/aiter#3832 gfx950 MXFP4 MoE backend. - image: rocm/vllm-dev:nightly_cdna4 + image: rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 model: amd/Kimi-K2.5-MXFP4 model-prefix: kimik2.5 runner: mi355x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 133df6a749..95c298d9ee 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5014,7 +5014,7 @@ - config-keys: - kimik2.5-fp4-mi355x-vllm description: - - "Set the vLLM ROCm image to rocm/vllm-dev:nightly_cdna4, which ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" + - "Set the vLLM ROCm image to the immutable rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 build (the nightly_cdna4 pin), which ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, and the AITER BF16 GEMM path" - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 From 19100b1036e9a063dc3a5ede0c3a9393a2ad7608 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Wed, 22 Jul 2026 02:38:51 +0000 Subject: [PATCH 6/9] fix(kimik2.5): lower MI355X gpu-memory-utilization 0.90 -> 0.85 MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit TP4/TP8 runs OOM during CUDA-graph capture: MI355X nodes hold ~30 GiB per GPU outside the vLLM process, so a 0.90 budget (259 GiB) leaves 0 bytes physically free and capture buffers cannot allocate. 0.85 keeps weights + KV inside 288 - 30 GiB with headroom for graph capture. 中文：将 MI355X 的 gpu-memory-utilization 从 0.90 下调到 0.85。TP4/TP8 在 CUDA graph 捕获阶段发生显存溢出——MI355X 节点每张卡有约 30 GiB 被 vLLM 进程之外占用，0.90 的预算（259 GiB）导致物理显存无剩余，捕获缓冲无法 分配。0.85 使权重与 KV 缓存控制在 288 - 30 GiB 内，为 graph 捕获留出余量。 Co-authored-by: Cursor --- benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh index 6dadc26876..ecd3d0e67a 100755 --- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh +++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_mi355x.sh @@ -76,7 +76,7 @@ set -x vllm serve $MODEL --port $PORT \ --tensor-parallel-size=$TP \ $EP \ ---gpu-memory-utilization 0.90 \ +--gpu-memory-utilization 0.85 \ --max-model-len $MAX_MODEL_LEN \ --kv-cache-dtype fp8 \ --block-size 16 \ From ca3845f2c7f6ba29d446a001135006bec64751a8 Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Wed, 22 Jul 2026 19:36:23 +0000 Subject: [PATCH 7/9] perf(amd): track floating nightly_cdna4 tag for Kimi-K2.5 MXFP4 MI355X MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The AITER MXFP4 MoE bump (vllm-project/vllm#48683 / ROCm/aiter#3832) has landed in the floating rocm/vllm-dev:nightly_cdna4 tag, so drop the immutable dated pin and track nightly_cdna4 directly. 中文：改为跟踪浮动的 rocm/vllm-dev:nightly_cdna4 标签。AITER MXFP4 MoE 的版本提升（vllm-project/vllm#48683 / ROCm/aiter#3832）已合入 nightly_cdna4 浮动标签，因此去掉固定日期的镜像 pin，直接跟踪 nightly_cdna4。 Co-authored-by: Cursor --- configs/amd-master.yaml | 2 +- perf-changelog.yaml | 4 ++-- 2 files changed, 3 insertions(+), 3 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index f0eeb3b849..a57bfc1af5 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -509,7 +509,7 @@ kimik2.5-int4-mi300x-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-mi355x-vllm: - image: rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 + image: rocm/vllm-dev:nightly_cdna4 model: amd/Kimi-K2.5-MXFP4 model-prefix: kimik2.5 runner: mi355x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 2907a7d374..254acabdf6 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5014,8 +5014,8 @@ - config-keys: - kimik2.5-fp4-mi355x-vllm description: - - "Set the vLLM ROCm image to the immutable rocm/vllm-dev:nightly_cdna4_main_torch_2.11.0_rocm7.14.0a20260623_0721_b101 build (the nightly_cdna4 pin), which ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" - - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, and the AITER BF16 GEMM path" + - "Set the vLLM ROCm image to the rocm/vllm-dev:nightly_cdna4 tag, which now ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" + - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 From 7efe2986c714f7642abe80e9ff0f1884d50a413a Mon Sep 17 00:00:00 2001 From: Fangzhou Ai <31551580+Fangzhou-Ai@users.noreply.github.com> Date: Wed, 22 Jul 2026 19:41:24 +0000 Subject: [PATCH 8/9] perf(amd): pin official vllm-openai-rocm nightly for Kimi-K2.5 MXFP4 MI355X MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit Switch the kimik2.5-fp4-mi355x-vllm image from the rocm/vllm-dev CDNA4 nightly to the official vllm/vllm-openai-rocm nightly, commit-pinned to the immutable nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 tag (latest as of 2026-07-22). It carries the AITER MXFP4 MoE bump (vllm-project/vllm#48683 / ROCm/aiter#3832) while staying reproducible. 中文：将 kimik2.5-fp4-mi355x-vllm 的镜像从 rocm/vllm-dev CDNA4 nightly 切换为官方 vllm/vllm-openai-rocm nightly，并按提交固定到不可变的 nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 标签（截至 2026-07-22 最新）。该镜像已包含 AITER MXFP4 MoE 的版本提升 （vllm-project/vllm#48683 / ROCm/aiter#3832），同时保持可复现。 Co-authored-by: Cursor --- configs/amd-master.yaml | 2 +- perf-changelog.yaml | 2 +- 2 files changed, 2 insertions(+), 2 deletions(-) diff --git a/configs/amd-master.yaml b/configs/amd-master.yaml index a57bfc1af5..8ea211a5de 100644 --- a/configs/amd-master.yaml +++ b/configs/amd-master.yaml @@ -509,7 +509,7 @@ kimik2.5-int4-mi300x-vllm: - { tp: 8, conc-start: 4, conc-end: 64 } kimik2.5-fp4-mi355x-vllm: - image: rocm/vllm-dev:nightly_cdna4 + image: vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 model: amd/Kimi-K2.5-MXFP4 model-prefix: kimik2.5 runner: mi355x diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 254acabdf6..686ed2e8ec 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5014,7 +5014,7 @@ - config-keys: - kimik2.5-fp4-mi355x-vllm description: - - "Set the vLLM ROCm image to the rocm/vllm-dev:nightly_cdna4 tag, which now ships AITER 0.1.19.dev (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" + - "Set the vLLM ROCm image to the immutable vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 build (the latest official ROCm nightly, commit-pinned), which ships the AITER MXFP4 MoE bump (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 From 5ef6877d726b0bf1dccb7495c757fdcc9b369c6c Mon Sep 17 00:00:00 2001 From: functionstackx <47992694+functionstackx@users.noreply.github.com> Date: Thu, 30 Jul 2026 17:58:23 -0400 Subject: [PATCH 9/9] fix(changelog): move the Kimi-K2.5 entry out of PR #2104's entry [skip-sweep] MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit The entry was inserted directly after another entry's `- config-keys:` line, so YAML folded `pr-link: .../2213` together with the following ` - minimaxm3-fp4-mi355x-vllm` into one scalar and the duplicate `description:`/`pr-link:` keys overwrote this PR's own values. Loading the file produced a single entry with config-keys [kimik2.5-fp4-mi355x-vllm] but PR #2104's description and pr-link, and #2104's entry vanished as a separate entry. check-changelog caught it: process_changelog.py parses only the added lines and they began mid-structure ("expected ''"). Move the six lines to the end of the file as their own entry, per the append-only convention, and restore #2104's entry. Net diff vs main is additions only; config-keys are unchanged, so sweep selection is unaffected and the PR's existing green sweep remains representative. 中文：该条目被插入到另一条目的 `- config-keys:` 之后，导致 YAML 将 `pr-link: .../2213` 与随后的 ` - minimaxm3-fp4-mi355x-vllm` 合并为同一标量，且重复 的 `description:`/`pr-link:` 键覆盖了本 PR 自身的取值：解析结果是 config-keys 为 kimik2.5 但描述与 pr-link 却来自 #2104，而 #2104 的条目不再独立存在。现按追加式约定 将这 6 行移至文件末尾成为独立条目，并恢复 #2104 的条目。相对 main 仅有新增、无删除； config-keys 未变，扫描选择不受影响，PR 既有的绿色扫描结果仍然适用。 Co-Authored-By: Claude Opus 5 (1M context) --- perf-changelog.yaml | 14 ++++++++------ 1 file changed, 8 insertions(+), 6 deletions(-) diff --git a/perf-changelog.yaml b/perf-changelog.yaml index 8cbf67c953..6eebd97070 100644 --- a/perf-changelog.yaml +++ b/perf-changelog.yaml @@ -5033,12 +5033,6 @@ pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2258 - config-keys: - - kimik2.5-fp4-mi355x-vllm - description: - - "Set the vLLM ROCm image to the immutable vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 build (the latest official ROCm nightly, commit-pinned), which ships the AITER MXFP4 MoE bump (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" - - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" - - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" - pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213 - minimaxm3-fp4-mi355x-vllm description: - "Close the high-concurrency gap vs the ATOM recipe on MiniMax-M3 MXFP4 MI355X single-node vLLM STP. Bump image to latest nightly vllm/vllm-openai-rocm:nightly-69715823df89b11ee684b84066390cbb9092d5c1." @@ -5353,3 +5347,11 @@ - "Add seven checked-in srt-slurm recipes selected one-to-one by the 8K/1K prefill/decode topology and concurrency entries." - "Image: vllm/vllm-openai:v0.25.1" pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2360 + +- config-keys: + - kimik2.5-fp4-mi355x-vllm + description: + - "Set the vLLM ROCm image to the immutable vllm/vllm-openai-rocm:nightly-387189c42997b27e2c04b5d97ef8190ffa2bf909 build (the latest official ROCm nightly, commit-pinned), which ships the AITER MXFP4 MoE bump (>= v0.1.16.post5 from vllm-project/vllm#48683) including the ROCm/aiter#3832 gfx950 MXFP4 MoE backend" + - "Apply the accuracy-gated Kimi-K2.5 MXFP4 settings: tuned AITER MXFP4 MoE, fused shared experts, FP8 KV cache, block size 16, 16384 batched tokens, 512 sequences, async scheduling, gpu-memory-utilization 0.85 (headroom for CUDA-graph capture on MI355X), and the AITER BF16 GEMM path" + - "Extend the TP4 and TP8 8k1k concurrency sweep from 64 to 128 (1k1k deprecated per #2263)" + pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2213