From 22f4d7f2eb9751d2c966fe443da38a83c2a8b4dc Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 17:40:25 -0400
Subject: [PATCH 01/11] kimik2.5 fp4 b300: v0.26.0 image, TP/DEP sweep, tuning

---
 .../fixed_seq_len/kimik2.5_fp4_b300.sh        | 28 +++++++++++++++++--
 configs/nvidia-master.yaml                    |  7 +++--
 perf-changelog.yaml                           |  6 ++++
 runners/launch_b300-nv.sh                     |  2 ++
 4 files changed, 38 insertions(+), 5 deletions(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index db6d3fb0d1..7cb3fc3d7e 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -16,6 +16,21 @@ check_env_vars \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME
 
+PARALLEL_ARGS=(--tensor-parallel-size "$TP" --data-parallel-size 1)
+GMU=0.90
+PREFILL_SCHEDULE_ARGS=()
+if [ "${DP_ATTENTION:-false}" = "true" ]; then
+    PARALLEL_ARGS=(--tensor-parallel-size 1 --data-parallel-size "$TP")
+    GMU=0.85
+    PREFILL_SCHEDULE_ARGS=(--prefill-schedule-interval 4)
+fi
+
+EP_ARGS=()
+if [ "${EP_SIZE:-1}" -gt 1 ]; then
+    EP_ARGS=(--enable-expert-parallel)
+fi
+
+
 # `hf download` creates the target dir if missing and is itself idempotent. 
 # When MODEL_PATH is unset (stand-alone runs), fall back to the HF_HUB_CACHE
 # Either way, MODEL_PATH is what the server is launched with.
@@ -37,6 +52,8 @@ nvidia-smi
 
 export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
+export VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS=""
+export VLLM_RPC_TIMEOUT=600000
 
 SERVER_LOG=/workspace/server.log
 
@@ -49,13 +66,20 @@ start_gpu_monitor
 
 set -x
 vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
---tensor-parallel-size $TP \
---gpu-memory-utilization 0.90 \
+"${PARALLEL_ARGS[@]}" \
+"${EP_ARGS[@]}" \
+"${PREFILL_SCHEDULE_ARGS[@]}" \
+--gpu-memory-utilization "$GMU" \
 --max-model-len $MAX_MODEL_LEN \
 --max-num-seqs $CONC \
 --reasoning-parser kimi_k2 \
 --tool-call-parser kimi_k2 \
 --compilation_config.pass_config.fuse_allreduce_rms true \
+--kv-cache-dtype fp8 \
+--max-cudagraph-capture-size "$CONC" \
+--stream-interval 32 \
+--attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
+--linear-backend flashinfer_cutedsl \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index a30af59641..3f2f2eb2b7 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1528,7 +1528,7 @@ kimik2.5-fp4-b200-vllm:
 # Kimi-K2.5 FP4 B200 vLLM recipe as-is until B300-specific tuning is available.
 
 kimik2.5-fp4-b300-vllm:
-  image: vllm/vllm-openai:v0.22.0
+  image: vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7
   model: nvidia/Kimi-K2.5-NVFP4
   model-prefix: kimik2.5
   runner: b300
@@ -1540,8 +1540,9 @@ kimik2.5-fp4-b300-vllm:
     - isl: 8192
       osl: 1024
       search-space:
-      - { tp: 8, ep: 1, conc-start: 1, conc-end: 4 }
-      - { tp: 4, ep: 1, conc-start: 1, conc-end: 128 }
+      - { tp: 8, ep: 1, conc-list: [1] }
+      - { tp: 4, ep: 1, conc-start: 1, conc-end: 512 }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512 }
 
 dsr1-fp8-b200-sglang-mtp:
   image: lmsysorg/sglang:v0.5.12-cu130
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 6eebd97070..274c65c56c 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5355,3 +5355,9 @@
     - "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
+
+- config-keys:
+    - kimik2.5-fp4-b300-vllm
+  description:
+    - "Kimi K2.5 NVFP4 B300 vLLM: v0.26.0 image, TP8 conc-1 only, TP4/DEP sweep, DEP gmu 0.85, tuning"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2442
diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index f39616801b..e22cf5d556 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -56,6 +56,8 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && $FRAMEWORK == "dy
 elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" && $FRAMEWORK == "dynamo-vllm" ]]; then
     export MODEL_PATH="/data/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
+elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
+    export MODEL_PATH="/data/models/Kimi-K2.5-NVFP4"
 else
     echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm or dynamo-sglang, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm"
     exit 1

From dfa95918a704cc036d0cd3bd519ba70ad20165ce Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 17:48:25 -0400
Subject: [PATCH 02/11] kimik2.5 fp4 b300: v0.26.0 image, TP/DEP sweep, tuning

---
 runners/launch_b300-nv.sh | 6 +++---
 1 file changed, 3 insertions(+), 3 deletions(-)

diff --git a/runners/launch_b300-nv.sh b/runners/launch_b300-nv.sh
index e22cf5d556..fdf9df08c0 100644
--- a/runners/launch_b300-nv.sh
+++ b/runners/launch_b300-nv.sh
@@ -56,8 +56,6 @@ elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp4" && $FRAMEWORK == "dy
 elif [[ $MODEL_PREFIX == "minimaxm3" && $PRECISION == "fp8" && $FRAMEWORK == "dynamo-vllm" ]]; then
     export MODEL_PATH="/data/models/MiniMax-M3-MXFP8"
     export SRT_SLURM_MODEL_PREFIX="MiniMaxAI/MiniMax-M3-MXFP8"
-elif [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
-    export MODEL_PATH="/data/models/Kimi-K2.5-NVFP4"
 else
     echo "Unsupported model: $MODEL_PREFIX-$PRECISION. Supported models are: dsr1-fp4, dsr1-fp8, dsv4-fp4 with dynamo-vllm or dynamo-sglang, minimaxm2.5-fp4 with dynamo-vllm, minimaxm2.5-fp8 with dynamo-vllm, minimaxm3-fp4 with dynamo-vllm, minimaxm3-fp8 with dynamo-vllm"
     exit 1
@@ -408,7 +406,9 @@ else
     # MODEL stays as the HF id for the client (--served-model-name, tokenizer);
     # MODEL_PATH is what the server reads weights from.
     MODEL_BASENAME="${MODEL##*/}"
-    if [[ " ${STAGED_MODELS[*]} " == *" ${MODEL_BASENAME} "* ]]; then
+    if [[ $MODEL_PREFIX == "kimik2.5" && $PRECISION == "fp4" ]]; then
+        export MODEL_PATH="${WRITABLE_MODELS_DIR%/}/${MODEL_BASENAME}"
+    elif [[ " ${STAGED_MODELS[*]} " == *" ${MODEL_BASENAME} "* ]]; then
         export MODEL_PATH="${HF_HUB_CACHE_MOUNT%/}/${MODEL_BASENAME}"
     else
         export MODEL_PATH="${WRITABLE_MODELS_DIR%/}/${MODEL_BASENAME}"

From 7746671708acf82ab8168689662f48fb1cff78c7 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 19:21:29 -0400
Subject: [PATCH 03/11] kimik2.5 fp4 b300: nightly image, extend sweep space
 with TP/DEP arms

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 1 -
 perf-changelog.yaml                                       | 2 +-
 2 files changed, 1 insertion(+), 2 deletions(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index 7cb3fc3d7e..a87fdd9587 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -79,7 +79,6 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --max-cudagraph-capture-size "$CONC" \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
---linear-backend flashinfer_cutedsl \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 
diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 274c65c56c..d8b4ef4955 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5359,5 +5359,5 @@
 - config-keys:
     - kimik2.5-fp4-b300-vllm
   description:
-    - "Kimi K2.5 NVFP4 B300 vLLM: v0.26.0 image, TP8 conc-1 only, TP4/DEP sweep, DEP gmu 0.85, tuning"
+    - "Kimi K2.5 NVFP4 B300 vLLM: nightly image, extend sweep space with TP/DEP arms, TP8 conc-1 only, DEP gmu 0.85"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2442

From 129b9b138e27fae51fb71ee300469c04ddbd87ce Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Sun, 2 Aug 2026 23:34:15 -0400
Subject: [PATCH 04/11] drop --attention-config from B300

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 1 -
 1 file changed, 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index a87fdd9587..d7e9fa2436 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -78,7 +78,6 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --kv-cache-dtype fp8 \
 --max-cudagraph-capture-size "$CONC" \
 --stream-interval 32 \
---attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

From 80d9ed859b529a75cb55de20c997ed004f7f97a7 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Mon, 3 Aug 2026 00:42:58 -0400
Subject: [PATCH 05/11] add attention-config with prefill query quant and
 linear-backend flashinfer_cutlass

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 2 ++
 1 file changed, 2 insertions(+)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index d7e9fa2436..59b67f316d 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -78,6 +78,8 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --kv-cache-dtype fp8 \
 --max-cudagraph-capture-size "$CONC" \
 --stream-interval 32 \
+--attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
+--linear-backend flashinfer_cutlass \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

From 20a5a284959edc58edf8d358fb2cb2be3dd87205 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Mon, 3 Aug 2026 00:45:56 -0400
Subject: [PATCH 06/11] set max-cudagraph-capture-size to 2x concurrency

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index 59b67f316d..fd93c88a47 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -76,7 +76,7 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --tool-call-parser kimi_k2 \
 --compilation_config.pass_config.fuse_allreduce_rms true \
 --kv-cache-dtype fp8 \
---max-cudagraph-capture-size "$CONC" \
+--max-cudagraph-capture-size "$((CONC * 2))" \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --linear-backend flashinfer_cutlass \

From 67892db888d3ca567d623480f48411f9947ee20a Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Mon, 3 Aug 2026 01:50:32 -0400
Subject: [PATCH 07/11] set max-num-batched-tokens to 16k

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index fd93c88a47..dba41ccee3 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -80,6 +80,7 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --linear-backend flashinfer_cutlass \
+--max-num-batched-tokens 16384 \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

From b13329c7c26500aecf1ee0cb264e0e367ba0b2be Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Mon, 3 Aug 2026 03:29:09 -0400
Subject: [PATCH 08/11] remove max-num-batched-tokens, set
 VLLM_USE_V2_MODEL_RUNNER=0

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index dba41ccee3..8f1906aaa3 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -52,6 +52,7 @@ nvidia-smi
 
 export TORCH_CUDA_ARCH_LIST="10.0"
 export PYTHONNOUSERSITE=1
+export VLLM_USE_V2_MODEL_RUNNER=0
 export VLLM_FLASHINFER_AUTOTUNE_SKIP_OPS=""
 export VLLM_RPC_TIMEOUT=600000
 
@@ -80,7 +81,6 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --linear-backend flashinfer_cutlass \
---max-num-batched-tokens 16384 \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

From aa76f274912afa8c119a259f075718e93ab88982 Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Tue, 4 Aug 2026 19:20:17 -0400
Subject: [PATCH 09/11] add --no-enable-flashinfer-autotune

---
 benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh | 1 +
 1 file changed, 1 insertion(+)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index 8f1906aaa3..f3ff0ea9c6 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -81,6 +81,7 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --linear-backend flashinfer_cutlass \
+--no-enable-flashinfer-autotune \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

From c1354604a90c4115f2de4be311c9203bbd8568bc Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Tue, 4 Aug 2026 19:22:05 -0400
Subject: [PATCH 10/11] fix perf-changelog

---
 perf-changelog.yaml | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

diff --git a/perf-changelog.yaml b/perf-changelog.yaml
index 01af97c83c..479597ad69 100644
--- a/perf-changelog.yaml
+++ b/perf-changelog.yaml
@@ -5450,12 +5450,8 @@
     - "Cap the 1P1D TP4 concurrency sweep at 256 (was 512); drop the 2P1D TP4 layout (128/256/512) as it is CI-flaky with negligible curve impact."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1943
 
-
 - config-keys:
-    - minimaxm3-fp4-mi355x-vllm-disagg
+    - kimik2.5-fp4-b300-vllm
   description:
-    - "Update the MiniMax-M3 MXFP4 MI355X vLLM disagg image to vllm/vllm-openai-rocm:nightly-2dfaae752b4db0d43cfc0715c780e33be030d0f1 (from rocm/vllm-dev:vllm-0.23.1-rocm723-mi35x-mori-0625) for AITER MoE and shared-expert fusion support."
-    - "Export VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1 for MiniMax-M3-MXFP4 (both prefill and decode)."
-    - "Enable prefill-only INT4 quick-reduce: set VLLM_ROCM_QUICK_REDUCE_QUANTIZATION=INT4 and VLLM_ROCM_QUICK_REDUCE_MAX_SIZE_BYTES_MB=2048 on the prefill workers via a new prefill_env channel (mirrors the existing decode_env path in server_vllm.sh)."
-    - "Cap the 1P1D TP4 concurrency sweep at 256 (was 512); drop the 2P1D TP4 layout (128/256/512) as it is CI-flaky with negligible curve impact."
-  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1943
+    - "Kimi K2.5 NVFP4 B300 vLLM: nightly image, extend sweep space with TP/DEP arms, TP8 conc-1 only, DEP gmu 0.85"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2442

From 53081ea1a443b257d8597b2e5149393335e37c4d Mon Sep 17 00:00:00 2001
From: Xin Li <xinli@nvidia.com>
Date: Wed, 5 Aug 2026 00:08:33 -0400
Subject: [PATCH 11/11] remove --no-enable-flashinfer-autotune, add TEP arms

---
 .../fixed_seq_len/kimik2.5_fp4_b300.sh         |  1 -
 configs/nvidia-master.yaml                     | 18 ++++++++++++++++++
 2 files changed, 18 insertions(+), 1 deletion(-)

diff --git a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
index f3ff0ea9c6..8f1906aaa3 100755
--- a/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
+++ b/benchmarks/single_node/fixed_seq_len/kimik2.5_fp4_b300.sh
@@ -81,7 +81,6 @@ vllm serve $MODEL_PATH --served-model-name $MODEL --host 0.0.0.0 --port $PORT \
 --stream-interval 32 \
 --attention-config '{"mla_prefill_backend":"FLASHINFER","use_prefill_query_quantization":true}' \
 --linear-backend flashinfer_cutlass \
---no-enable-flashinfer-autotune \
 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 
diff --git a/configs/nvidia-master.yaml b/configs/nvidia-master.yaml
index 22f4c190cc..1e9f93cbd8 100644
--- a/configs/nvidia-master.yaml
+++ b/configs/nvidia-master.yaml
@@ -1436,6 +1436,24 @@ kimik2.5-fp4-b200-vllm:
 # Kimi-K2.5 FP4 B200 vLLM recipe as-is until B300-specific tuning is available.
 
 kimik2.5-fp4-b300-vllm:
+  image: vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7
+  model: nvidia/Kimi-K2.5-NVFP4
+  model-prefix: kimik2.5
+  runner: b300
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    fixed-seq-len:
+    - isl: 8192
+      osl: 1024
+      search-space:
+      - { tp: 8, ep: 1, conc-list: [1] }
+      - { tp: 4, ep: 1, conc-start: 1, conc-end: 512 }
+      - { tp: 8, ep: 8, dp-attn: false, conc-list: [1] }
+      - { tp: 4, ep: 4, dp-attn: false, conc-start: 1, conc-end: 512 }
+      - { tp: 4, ep: 4, dp-attn: true, conc-start: 128, conc-end: 512 }
+-vllm:
   image: vllm/vllm-openai:nightly-e2fa28594f7baad142a426b0b6a2cfe2c79201c7
   model: nvidia/Kimi-K2.5-NVFP4
   model-prefix: kimik2.5