From 743b7618d0f6878029c6abb550d878579e0dd38e Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 3 Aug 2026 11:27:57 -0500
Subject: [PATCH 01/10] perf(agentx): refresh B300 Kimi K3 DSpark

---
 perf-changelog.yaml | 5 +++++
 1 file changed, 5 insertions(+)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 7a064f8090..4bdc05c19c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5374,3 +5374,8 @@
     - "Image: lmsysorg/sglang:v0.5.16-cu130"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2420
   
+- config-keys:
+    - kimik3-fp4-b300-vllm-agentic-dspark
+  description:
+    - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/0

From 1103b865c88d7c9ffd59cf3ce0892105d07c2928 Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Mon, 3 Aug 2026 11:28:22 -0500
Subject: [PATCH 02/10] chore: link B300 DSpark refresh PR

---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 4bdc05c19c..f875f837b7 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5378,4 +5378,4 @@
     - kimik3-fp4-b300-vllm-agentic-dspark
   description:
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/0
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476

From 98b563c0fb33276566a6fd5a7bd7860e7971028c Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 10:16:01 -0500
Subject: [PATCH 03/10] fix: test B300 with modern Kimi nightly
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：使用新版 Kimi nightly 镜像测试 B300
---
 configs/nvidia-master.yaml | 2 +-
 perf-changelog.yaml        | 1 +
 2 files changed, 2 insertions(+), 1 deletion(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 393babf91b..0359ea8770 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1521,7 +1521,7 @@ kimik3-fp4-b300-vllm-agentic-dspark:
   # its native window is 32k YaRN-stretched to 1M, and the same image/draft/config
   # reaches AL 4.18 on short-context work. These throughput numbers measure the
   # system at a prescribed acceptance, not the draft's fitness at 100k+ context.
-  image: vllm/vllm-openai:kimi-k3
+  image: vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:b300-nv
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 30cb75861a..bd45c8a378 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5715,4 +5715,5 @@
     - kimik3-fp4-b300-vllm-agentic-dspark
   description:
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
+    - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476
\ No newline at end of file

From dbf99d878b4eca8215d04c3a41754bdb21db71fe Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 11:10:58 -0500
Subject: [PATCH 04/10] fix: terminate Kimi changelog entry
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：补齐 Kimi 基准测试变更日志条目的末尾换行。
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bd45c8a378..bfb1d86f7f 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5716,4 +5716,4 @@
   description:
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
     - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476
\ No newline at end of file
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476

From c86cb7e4072235ec8cab6bf7e1ac0e415917b640 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 11:31:27 -0500
Subject: [PATCH 05/10] fix: cap Kimi resident concurrency
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 Kimi GPU 常驻基准测试的并发上限设为 8，并保留 DRAM 卸载在并发 16 的容量测试点。
---
 configs/nvidia-master.yaml | 18 ++++++++----------
 perf-changelog.yaml        |  1 +
 2 files changed, 9 insertions(+), 10 deletions(-)

diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 0359ea8770..ae4a509bd1 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1533,17 +1533,15 @@ kimik3-fp4-b300-vllm-agentic-dspark:
     agentic-coding:
     - dram-utilization: 0.63
       search-space:
-      # Mirrors the non-MTP entry's KV arms so the spec-decoding delta is
-      # readable at equal concurrency, but stops at conc 16 rather than 24: the
-      # non-MTP bring-up sweep (run 30326393603) showed the GPU-resident arm
-      # already thrashing at conc >= 16 (prefix cache hit 2.7%, TTFT p50 86-191s)
-      # because GPU KV holds only ~3.1 max-length requests, so conc 24 costs a
-      # full job per arm to re-measure that regime. conc 16 still exercises the
-      # DRAM tier meaningfully (62% external prefix cache hit rate there).
-      # TP8-only for the same memory reason: a ~1.5 TB MXFP4 checkpoint does not
-      # fit below 8 GPUs.
+      # Keep equal resident/offload points through conc 8. The digest-pinned
+      # bring-up sweep (run 31404943911) confirmed that resident conc 16
+      # remained in warmup after 1,770 seconds (96/177 requests), while conc 8
+      # entered profiling. The offload conc 16 arm completed warmup in 804
+      # seconds and exercised the external KV tier, so retain it as the
+      # capacity endpoint.
+      # TP8-only: a ~1.5 TB MXFP4 checkpoint does not fit below 8 GPUs.
       # TP8 GPU-resident
-      - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8, 16] }
+      - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 2, 4, 8] }
       # TP8 SimpleCPUOffload (host DRAM)
       - { tp: 8, ep: 1, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [1, 2, 4, 8, 16] }
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index bfb1d86f7f..157a47a4e2 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5716,4 +5716,5 @@
   description:
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
     - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
+    - "Cap the GPU-resident search at concurrency 8 after concurrency 16 failed to complete deterministic warmup; retain concurrency 16 for the DRAM-offload capacity tier"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476

From ddc85dcce1120e97010f91cd9a0a03fa7d4c423f Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 19:04:45 -0500
Subject: [PATCH 06/10] fix: preserve changelog newline after merge
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：合并 main 后保留性能变更日志末尾换行
---
 perf-changelog.yaml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 863460dd78..ba48df5fb4 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5770,4 +5770,4 @@
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
     - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
     - "Cap the GPU-resident search at concurrency 8 after concurrency 16 failed to complete deterministic warmup; retain concurrency 16 for the DRAM-offload capacity tier"
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476
\ No newline at end of file
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476

From 77021db16a41c645cd057969b840e936723bea46 Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 19:32:09 -0500
Subject: [PATCH 07/10] ci: exclude failed B300 NVLink node
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：排除发生不可纠正 NVLink 错误的 B300 节点，避免基准测试继续调度到故障硬件。
---
 .github/workflows/benchmark-tmpl.yml | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/.github/workflows/benchmark-tmpl.yml b/.github/workflows/benchmark-tmpl.yml
index 6c4fe50fe5..257adcb1d9 100644
--- a/.github/workflows/benchmark-tmpl.yml
+++ b/.github/workflows/benchmark-tmpl.yml
@@ -197,7 +197,7 @@ env:
   MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
   MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
   # These b300 nodes are currently broken.
-  SALLOC_EXCLUDE: 'b300-005,b300-006'
+  SALLOC_EXCLUDE: 'b300-005,b300-006,b300-017'
 
 permissions:
   contents: read

From f7609f22f45445ca425ca954f48b57bf554f9fba Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Mon, 10 Aug 2026 19:56:49 -0500
Subject: [PATCH 08/10] ci: honor B300 node exclusions
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：让 B300 启动器把 SALLOC_EXCLUDE 传递给 Slurm，避免任务调度到已知故障节点。
---
 runners/launch_b300-nv.sh | 6 +++++-
 1 file changed, 5 insertions(+), 1 deletion(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index cad9ba30aa..29cb0b418a 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -482,7 +482,11 @@ else
     export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
 
     SALLOC_TIME_LIMIT="${SALLOC_TIME_LIMIT:-480}"
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT -N 1 --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
+    SALLOC_EXCLUDE_ARGS=()
+    if [[ -n "${SALLOC_EXCLUDE:-}" ]]; then
+        SALLOC_EXCLUDE_ARGS+=(--exclude="$SALLOC_EXCLUDE")
+    fi
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT -N 1 --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" "${SALLOC_EXCLUDE_ARGS[@]}" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID \

From 02ca07dd851dda2336ae391e4d45b2e5879396be Mon Sep 17 00:00:00 2001
From: adibarra <93070681+adibarra@users.noreply.github.com>
Date: Tue, 11 Aug 2026 10:49:29 -0500
Subject: [PATCH 09/10] fix: align kimi k3 b300 prefill
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

中文：将 Kimi K3 B300 的 MLA 预填充后端对齐到上游 FlashInfer 配置，并保留真实 block 验证的评估路径。
---
 .../single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh | 13 +++++++------
 perf-changelog.yaml                                 |  2 ++
 2 files changed, 9 insertions(+), 6 deletions(-)

diff --git a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
index faf2232726..080f63aea0 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
@@ -226,9 +226,10 @@ VLLM_CMD=(
     # --no-enable-flashinfer-autotune, VLLM_USE_V2_MODEL_RUNNER=1) killed the
     # engine core on every concurrency with an assertion in the FlashInfer MoE
     # runner's shared-experts output buffer
-    # (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once). Only
-    # mla_prefill_backend=TRTLLM_RAGGED is retained from that set; the rest stay
-    # on the values that ran 12/12 green in run 30326393603.
+    # (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once). Those
+    # unrelated values stay on the settings that ran 12/12 green in run
+    # 30326393603; the isolated MLA prefill choice below follows the current
+    # published Kimi-K3 FP8-KV profile.
     --gpu-memory-utilization 0.90
     --max-num-seqs "$MAX_NUM_SEQS"
     --max-model-len 1048576
@@ -240,9 +241,9 @@ VLLM_CMD=(
     --reasoning-parser kimi_k3
     --tool-call-parser kimi_k3
     --enable-auto-tool-choice
-    # FP8 KV cache requires the prefill query quantization flag. MLA prefill
-    # runs on TRTLLM_RAGGED per the upstream Blackwell override.
-    --attention-config '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
+    # FP8 KV cache requires prefill query quantization. Use FlashInfer MLA
+    # prefill, matching the current published upstream Kimi-K3 recipe.
+    --attention-config '{"mla_prefill_backend":"flashinfer","use_prefill_query_quantization":true}'
     --speculative-config "$SPEC_CONFIG"
     --compilation-config "$COMPILATION_CONFIG"
     --disable-uvicorn-access-log
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 639055164d..6586aff121 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5777,5 +5777,7 @@
   description:
     - "Refresh B300 Kimi K3 DSpark submission with corrected AgentX harness"
     - "Use the pinned 2026-08-10 vLLM nightly with upstream Kimi K3 and DSpark support"
+    - "Match the current published upstream Kimi K3 FP8-KV profile by using flashinfer MLA prefill with prefill query quantization"
+    - "Keep DSpark K=2 probabilistic throughput on synthetic golden AL 2.51 while the generated EVAL_ONLY row uses real block verification"
     - "Cap the GPU-resident search at concurrency 8 after concurrency 16 failed to complete deterministic warmup; retain concurrency 16 for the DRAM-offload capacity tier"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2476

From 5eaa7c6242334239f46f77f1294d94a7ac06338b Mon Sep 17 00:00:00 2001
From: Cam Quilici <cjquilici@gmail.com>
Date: Tue, 11 Aug 2026 16:07:20 -0500
Subject: [PATCH 10/10] chore: trim unrelated B300 changes
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Remove the PR-specific Slurm exclusion changes and restore the unrelated benchmark comment so the PR stays focused on the Kimi K3 refresh.\n\n中文：移除本 PR 中与 Slurm 节点排除相关的改动，并恢复无关的基准测试注释，使 PR 仅聚焦于 Kimi K3 刷新。
---
 .github/workflows/benchmark-tmpl.yml                       | 2 +-
 benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh | 7 +++----
 runners/launch_b300-nv.sh                                  | 6 +-----
 3 files changed, 5 insertions(+), 10 deletions(-)

diff --git a/.github/workflows/benchmark-tmpl.yml b/.github/workflows/benchmark-tmpl.yml
index 257adcb1d9..6c4fe50fe5 100644
--- a/.github/workflows/benchmark-tmpl.yml
+++ b/.github/workflows/benchmark-tmpl.yml
@@ -197,7 +197,7 @@ env:
   MODAL_TOKEN_ID: ${{ secrets.MODAL_TOKEN_ID }}
   MODAL_TOKEN_SECRET: ${{ secrets.MODAL_TOKEN_SECRET }}
   # These b300 nodes are currently broken.
-  SALLOC_EXCLUDE: 'b300-005,b300-006,b300-017'
+  SALLOC_EXCLUDE: 'b300-005,b300-006'
 
 permissions:
   contents: read
diff --git a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
index 080f63aea0..1d9bff1320 100755
--- a/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
+++ b/benchmarks/single_node/agentic/kimik3_fp4_b300_vllm_mtp.sh
@@ -226,10 +226,9 @@ VLLM_CMD=(
     # --no-enable-flashinfer-autotune, VLLM_USE_V2_MODEL_RUNNER=1) killed the
     # engine core on every concurrency with an assertion in the FlashInfer MoE
     # runner's shared-experts output buffer
-    # (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once). Those
-    # unrelated values stay on the settings that ran 12/12 green in run
-    # 30326393603; the isolated MLA prefill choice below follows the current
-    # published Kimi-K3 FP8-KV profile.
+    # (fused_moe/runner/shared_experts.py:165, all 8 TP ranks at once). Only
+    # mla_prefill_backend=TRTLLM_RAGGED is retained from that set; the rest stay
+    # on the values that ran 12/12 green in run 30326393603.
     --gpu-memory-utilization 0.90
     --max-num-seqs "$MAX_NUM_SEQS"
     --max-model-len 1048576
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index 29cb0b418a..cad9ba30aa 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -482,11 +482,7 @@ else
     export GPU_COUNT="${GPU_COUNT:-${TP:?TP must be set}}"
 
     SALLOC_TIME_LIMIT="${SALLOC_TIME_LIMIT:-480}"
-    SALLOC_EXCLUDE_ARGS=()
-    if [[ -n "${SALLOC_EXCLUDE:-}" ]]; then
-        SALLOC_EXCLUDE_ARGS+=(--exclude="$SALLOC_EXCLUDE")
-    fi
-    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT -N 1 --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" "${SALLOC_EXCLUDE_ARGS[@]}" --no-shell --job-name="$RUNNER_NAME"
+    salloc --partition=$SLURM_PARTITION --account=$SLURM_ACCOUNT -N 1 --gres=gpu:$GPU_COUNT --exclusive --mem=0 --time="$SALLOC_TIME_LIMIT" --no-shell --job-name="$RUNNER_NAME"
     JOB_ID=$(squeue --name="$RUNNER_NAME" -u "$USER" -h -o %A | head -n1)
 
     srun --jobid=$JOB_ID \