SemiAnalysisAI · wzhao18 · Sep 3, 2026 · Sep 3, 2026 · Sep 3, 2026 · Sep 3, 2026
diff --git a/...t-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark7-maxseq2-mooncake-agentic.yaml b/...t-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-dspark7-maxseq2-mooncake-agentic.yaml
@@ -1,19 +1,15 @@
-# p90-ITL Pareto family from OCI jobs 471335-471337 and 471341-471342: one
-# aggregate TP8 x DCP8 worker with max-num-seqs 2. Drafts at DSpark K=7 to
-# match the B300 low-latency arm (kimik3_fp4_b300_vllm_mtp.sh drafts at K=7
-# for conc <= 8) rather than the K=4 the measured jobs used.
 name: "kimi-k3-vllm-agg-gb300-dcp8-dspark7-maxseq2-mooncake-agentic"
 
 model:
   path: "moonshotai/Kimi-K3"
-  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   frameworks:
     dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
 
@@ -43,7 +39,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
-    router-mode: "random"
+    router-mode: "least-loaded"
     router-session-affinity-ttl-secs: 900
   env:
     DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
@@ -54,68 +50,59 @@ backend:
   connector: null
   mooncake_kv_store:
     master_extra_args:
+      - "--default_kv_lease_ttl=60000"
       - "--eviction_high_watermark_ratio=0.95"
       - "--eviction_ratio=0.10"
     store_config:
       metadata_server: "P2PHANDSHAKE"
-      global_segment_size: "150GB"
+      global_segment_size: "160GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
-      # Preserve the proven GB300 runner fabric rather than copying PR #211's
-      # single mlx5_0 rail, which is specific to the rail-isolated B300 fleet.
       device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
       mode: "embedded"
       enable_offload: false
   aggregated_environment:
     VLLM_ALLREDUCE_USE_FLASHINFER: "1"
     VLLM_ENABLE_K3_LATENT_MOE_TAIL_FUSION: "1"
-    VLLM_USE_V2_MODEL_RUNNER: "1"
-    # Use the direct DCP a2a/gather kernels rather than the collective
-    # fallback, matching the B300 arm (kimik3_fp4_b300_vllm_mtp.sh).
     VLLM_USE_DIRECT_DCP_A2A: "1"
     VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
     VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
+    ETCD_LEASE_TTL: "120"
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
+    VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_RUST_FRONTEND: "1"
+    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
     VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
     VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: "0"
-    PYTHONNOUSERSITE: "1"
-    TORCH_CUDA_ARCH_LIST: "10.0"
-    PYTHONHASHSEED: "42"
-    VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "900"
-    MC_GID_INDEX: "3"
-    MC_STORE_MEMCPY: "1"
-    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_SLICE_SIZE: "1048576"
-    MC_WORKERS_PER_CTX: "4"
     WITH_NVIDIA_PEERMEM: "0"
-    VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+    VLLM_LOG_STATS_INTERVAL: "1"
+
   vllm_config:
     aggregated:
+      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"compact_group_io":true,"max_load_batch_keys":2,"enable_offload":false}}'
       served-model-name: "moonshotai/Kimi-K3"
+      load-format: "fastsafetensors"
+      kv-cache-dtype: "fp8"
+      gpu-memory-utilization: 0.92
       tensor-parallel-size: 8
       decode-context-parallel-size: 8
       dcp-comm-backend: "a2a"
       max-num-seqs: 2
-      gpu-memory-utilization: 0.92
       max-num-batched-tokens: 8192
       trust-remote-code: true
+      max-cudagraph-capture-size: 1024
+      stream-interval: 10
       language-model-only: true
-      load-format: "fastsafetensors"
       moe-backend: "auto"
       no-enable-flashinfer-autotune: true
       enable-cumem-allocator: true
-      enable-prefix-caching: true
-      prefix-match-unit: 128
-      kv-cache-dtype: "fp8"
-      stream-interval: 10
       attention-backend: "TOKENSPEED_MLA"
       attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
-      # Throughput jobs inject the committed K=7 golden AL (3.84); EVAL_ONLY
-      # preserves this real target-verification configuration.
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":7,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
-      max-cudagraph-capture-size: 1024
-      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_offload":false}}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":7,"rejection_sample_method":"block"}'
+      enable-prefix-caching: true
+      prefix-match-unit: 128
 
 sbatch_directives:
   mem: "0"
@@ -132,10 +119,6 @@ benchmark:
   command: bash /infmax-workspace/benchmarks/multi_node/agentic_srt.sh
   env:
     INFMAX_CONTAINER_WORKSPACE: "/infmax-workspace"
-    AIPERF_TRACE_IDLE_GAP_CAP_SECONDS: "300"
-    AIPERF_LIVE_FAILED_REQUEST_THRESHOLD: "0.25"
-    AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
-    AIPERF_HTTP_TCP_USER_TIMEOUT: "900000"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"
     IS_MULTINODE: "true"

diff --git a/...i_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml b/...i_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-gb300-dcp8-nospec-mooncake-agentic.yaml
@@ -1,17 +1,15 @@
-# Measured p90-ITL Pareto family: OCI jobs 471604 (c32), 471605 (c48), and
-# 471606 (c64), using one aggregate TP8 x DCP8 worker without a draft model.
 name: "kimi-k3-vllm-agg-gb300-dcp8-nospec-mooncake-agentic"
 
 model:
   path: "moonshotai/Kimi-K3"
-  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   frameworks:
     dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
 
@@ -41,7 +39,7 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
-    router-mode: "random"
+    router-mode: "least-loaded"
     router-session-affinity-ttl-secs: 900
   env:
     DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
@@ -52,11 +50,12 @@ backend:
   connector: null
   mooncake_kv_store:
     master_extra_args:
+      - "--default_kv_lease_ttl=60000"
       - "--eviction_high_watermark_ratio=0.95"
       - "--eviction_ratio=0.10"
     store_config:
       metadata_server: "P2PHANDSHAKE"
-      global_segment_size: "150GB"
+      global_segment_size: "160GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
       device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
@@ -71,19 +70,17 @@ backend:
     VLLM_USE_DIRECT_DCP_A2A: "1"
     VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
     VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
+    ETCD_LEASE_TTL: "120"
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
+    VLLM_USE_RUST_FRONTEND: "1"
     VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
     VLLM_MEMORY_PROFILER_ESTIMATE_CUDAGRAPHS: "0"
     PYTHONNOUSERSITE: "1"
     TORCH_CUDA_ARCH_LIST: "10.0"
     PYTHONHASHSEED: "42"
     VLLM_HTTP_TIMEOUT_KEEP_ALIVE: "900"
-    MC_GID_INDEX: "3"
-    MC_STORE_MEMCPY: "1"
-    MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     MC_SLICE_SIZE: "1048576"
-    MC_WORKERS_PER_CTX: "4"
     WITH_NVIDIA_PEERMEM: "0"
     VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
   vllm_config:
@@ -106,7 +103,7 @@ backend:
       attention-backend: "TOKENSPEED_MLA"
       attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       max-cudagraph-capture-size: 1024
-      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_offload":false}}'
+      kv-transfer-config: '{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"compact_group_io":true,"max_load_batch_keys":2,"enable_offload":false}}'
 
 sbatch_directives:
   mem: "0"

diff --git a/...rm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml b/...rm-recipes/vllm/kimi-k3/agentic/disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic.yaml
@@ -1,18 +1,15 @@
-# Built vLLM image containing the Kimi-K3 DCP, DSpark, and PMU fixes.
-# Measured Pareto family: OCI jobs 469790 (c32), 469796 (c48), and 469792
-# (c64), using DSpark K=4 with synthetic acceptance length 3.36.
 name: "kimi-k3-vllm-disagg-gb300-1p1d-dcp8-dcp8-dspark4-mooncake-agentic"
 
 model:
   path: "moonshotai/Kimi-K3"
-  container: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+  container: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-dev-arm64-cu13.0.1-75c2eef"
+    image: "vllm/vllm-openai:nightly-dev-arm64-cu13-3696c77"
   frameworks:
     dynamo: "ba83080ecd31c1ce918559e576d3c5bc9e092ff1"
 
@@ -45,137 +42,106 @@ frontend:
   type: dynamo
   enable_multiple_frontends: false
   args:
-    router-mode: "random"
+    router-mode: "least-loaded"
     router-session-affinity-ttl-secs: 900
   env:
     DYN_ROUTER_ACTIVE_REQUEST_EXPIRY_SECS: "3600"
     DYN_TOKENIZER_CACHE_BYTES: "8589934592"
+
 backend:
   type: vllm
   connector: null
   dp_launch_mode: per_node
   kv_events_config:
     prefill: true
   mooncake_kv_store:
+    master_extra_args:
+      - "--default_kv_lease_ttl=60000"
+      - "--eviction_high_watermark_ratio=0.95"
+      - "--eviction_ratio=0.10"
     store_config:
       metadata_server: "P2PHANDSHAKE"
-      # Keep the colleague's measured 150 GB Mooncake segment setting.
-      global_segment_size: "150GB"
+      global_segment_size: "160GB"
       local_buffer_size: "4GB"
       protocol: "rdma"
       device_name: "mlx5_0,mlx5_1,mlx5_2,mlx5_3"
       mode: "embedded"
       enable_offload: false
   prefill_environment: &kimi_env
-    # Use the direct DCP a2a/gather kernels rather than the collective
-    # fallback, matching the B300 arm (kimik3_fp4_b300_vllm_mtp.sh).
     VLLM_USE_DIRECT_DCP_A2A: "1"
     VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
     VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
     VLLM_ALLREDUCE_USE_FLASHINFER: "1"
-    # No-op in this recipe: shared-expert sequence-parallel sharding requires
-    # expert parallelism, which is disabled on both roles.
     VLLM_KIMI_K3_SHARD_SP_SHARED_EXPERT: "0"
     DYN_REQUEST_PLANE: "tcp"
     ETCD_LEASE_TTL: "600"
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_RPC_TIMEOUT: "600000"
     TILELANG_CLEANUP_TEMP_FILES: "1"
     VLLM_USE_NCCL_SYMM_MEM: "0"
-    # TORCH_SYMMMEM deliberately NOT set to NVSHMEM here.
-    # MegaMoE (decode role) allocates its symm buffer with device='cuda' (no
-    # index, deep_gemm/mega/__init__.py:48) and calls symm_mem.rendezvous. On
-    # the NVSHMEM backend that reaches a process-global TeamManager singleton
-    # pinned to the first caller's device (nvshmem_team_manager.hpp:36-44), so a
-    # TP>1 engine whose workers span devices 0-3 dies with
-    #   "Detected use of TeamManager on multiple devices."
-    # Measured: fine at TP=1 (DEP16 decode, job 2649641), fails at TP>1 (job
-    # 2649921 prefill). Leaving it unset selects torch's default CUDA symm-mem
-    # backend, which has no such singleton.
     NCCL_CUMEM_ENABLE: "1"
     NCCL_MNNVL_ENABLE: "1"
     NCCL_NVLS_ENABLE: "1"
     VLLM_SERVER_DEV_MODE: "1"
     VLLM_USE_V2_MODEL_RUNNER: "1"
     VLLM_USE_RUST_FRONTEND: "1"
     VLLM_MOONCAKE_LOAD_RECV_THREADS: "4"
+    MC_SLICE_SIZE: "1048576"
     VLLM_CONNECTOR_PREFETCH_DEPTH: "8"
     VLLM_CONNECTOR_PREFETCH_KV_CAP: "0.65"
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
     VLLM_PREFIX_CACHE_RETENTION_INTERVAL: "0"
-    NCCL_P2P_LEVEL: NVL
+    NCCL_P2P_LEVEL: "NVL"
     MC_ENABLE_DEST_DEVICE_AFFINITY: "1"
     WITH_NVIDIA_PEERMEM: "0"
-    # NIXL (venv wheel) bundles its own UCX; NCCL's rdma_sharp_plugin loads
-    # /opt/hpcx UCX. Both in one process share UCM memory hooks and corrupt the
-    # heap -- NixlConnector then dies with
-    #   nixl_agent.cpp:470 registerMem: registration failed ... NIXL_ERR_BACKEND
-    # (and a "corrupted size vs. prev_size" abort at teardown). Disabling the
-    # external NCCL net plugin leaves exactly one UCX in the process; NCCL falls
-    # back to its built-in IB/NVLink transports. Verified on theia0263: 4 ranks,
-    # hpcx libs mapped = 0, 93-region registerMem OK on every rank.
     NCCL_NET_PLUGIN: "none"
-    # UCX_MEMTYPE_CACHE=n stops UCX misclassifying a CUDA pointer it did not
-    # allocate as host memory; REG_WHOLE=n keeps it from registering the whole
-    # enclosing region instead of the requested range.
     UCX_MEMTYPE_CACHE: "n"
     UCX_MEMTYPE_REG_WHOLE: "n"
     UCX_RCACHE_MAX_UNRELEASED: "1024"
     UCX_TCP_AF_PRIO: "inet"
-    # K3's linear-attention layers hold Mamba conv state. NixlConnector's
-    # 3-read conv transfer asserts the DS (dim, state_len) layout at worker
-    # init; without this both roles die with "requires DS conv state layout".
-    # Must match on prefill and decode, hence the shared anchor.
     VLLM_SSM_CONV_STATE_LAYOUT: "DS"
   decode_environment:
     <<: *kimi_env
 
   vllm_config:
     prefill:
-      # MultiConnector: Nixl (P->D handshake transfer) + Mooncake (external DRAM KV).
-      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"compact_group_io":true,"max_load_batch_keys":2,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
       served-model-name: "moonshotai/Kimi-K3"
-      enable-prefix-caching: true
       prefix-match-unit: 128
-      load-format: fastsafetensors
-      kv-cache-dtype: fp8
+      load-format: "fastsafetensors"
+      kv-cache-dtype: "fp8"
+      gpu-memory-utilization: 0.92
       tensor-parallel-size: 8
-      # The point of this recipe: DCP reuses the 8 TP ranks and shards the MLA
-      # KV cache across them instead of replicating it, giving the prefill role
-      # ~7.5x the KV pool it has under plain TP8.
       decode-context-parallel-size: 8
+      dcp-comm-backend: "a2a"
       enable-cumem-allocator: true
       trust-remote-code: true
-      no-enable-flashinfer-autotune: true
       max-cudagraph-capture-size: 512
       stream-interval: 10
-      max_num_batched_tokens: 16384
       language-model-only: true
       attention-backend: "TOKENSPEED_MLA"
       attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"rejection_sample_method":"block"}'
+      enable-prefix-caching: true
     decode:
-      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
+      kv-transfer-config: '{"kv_connector":"MultiConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"connectors":[{"kv_connector":"NixlConnector","kv_role":"kv_both","kv_load_failure_policy":"fail","kv_buffer_device":"cuda","kv_connector_extra_config":{"enforce_handshake_compat":false,"enable_cross_layers_blocks":false}},{"kv_connector":"MooncakeStoreConnector","kv_role":"kv_both","kv_load_failure_policy":"recompute","kv_connector_extra_config":{"load_async":true,"lookup_async":true,"compact_group_io":true,"max_load_batch_keys":2,"enable_cross_layers_blocks":false,"enable_offload":false}}]}}'
       served-model-name: "moonshotai/Kimi-K3"
-      enable-prefix-caching: true
       prefix-match-unit: 128
-      load-format: fastsafetensors
-      kv-cache-dtype: fp8
-      # Must equal the prefill TP for the NIXL MLA rank-to-rank mapping (see
-      # the header); DCP must likewise match at 8.
+      load-format: "fastsafetensors"
+      kv-cache-dtype: "fp8"
+      gpu-memory-utilization: 0.92
       tensor-parallel-size: 8
       decode-context-parallel-size: 8
+      dcp-comm-backend: "a2a"
       enable-cumem-allocator: true
       trust-remote-code: true
-      no-enable-flashinfer-autotune: true
-      compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
-      max-cudagraph-capture-size: 1024
+      max-cudagraph-capture-size: 512
       stream-interval: 10
-      max-num-seqs: 256
       language-model-only: true
       attention-backend: "TOKENSPEED_MLA"
       attention-config: '{"mla_prefill_backend": "TRTLLM_RAGGED", "use_prefill_query_quantization": true}'
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","attention_backend":"TOKENSPEED_MLA","method":"dspark","num_speculative_tokens":4,"rejection_sample_method":"block"}'
+      enable-prefix-caching: true
 
 sbatch_directives:
   mem: "0"
@@ -201,5 +167,3 @@ benchmark:
     AIPERF_DATASET_MMAP_CACHE_DIR: "/aiperf_mmap_cache"
     HF_HUB_CACHE: "/hf_hub_cache"
     WEKA_LOADER_OVERRIDE: "semianalysis_cc_traces_weka_062126"
-    # MODEL, topology, concurrency, duration, offload metadata, and result name
-    # are inherited from the generated InferenceMAX matrix.