[wip][nv] k3 b200 dcp8

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml

@@ -6,14 +6,14 @@ name: "kimik3-vllm-agg-b200-tp8dp2-latency-dspark-agentic"
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=b200&nodes=2&strategy=multi_node_dep
 model:
   path: "kimik3"
-  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+    image: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
 
 dynamo:
   install: false
@@ -50,6 +50,9 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
     VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_DIRECT_DCP_A2A: "1"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
     VLLM_USE_RUST_FRONTEND: "1"
     VLLM_SERVER_DEV_MODE: "1"
     VLLM_LOG_STATS_INTERVAL: "1"
@@ -65,6 +68,8 @@ backend:
       served-model-name: "moonshotai/Kimi-K3"
       tensor-parallel-size: 8
       data-parallel-size: 2
+      decode-context-parallel-size: 8
+      dcp-comm-backend: a2a
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: fastsafetensors
@@ -74,18 +79,23 @@ backend:
       no-enable-flashinfer-autotune: true
       max-model-len: 1048576
       kv-cache-dtype: fp8
-      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      attention-backend: TOKENSPEED_MLA
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       enable-prefix-caching: true
       enable-prompt-tokens-details: true
       enable-auto-tool-choice: true
       tool-call-parser: kimi_k3
       reasoning-parser: kimi_k3
       language-model-only: true
-      max-num-seqs: 8
-      max-num-batched-tokens: 8192
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      prefix-match-unit: 128
+      stream-interval: 10
+      enable-cumem-allocator: true
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[8,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,256,512,1024,2048,4096,8192]}'
       # Throughput runs use the committed K7 golden synthetic AL 3.84. Eval
       # selects the paired real-verification recipe instead.
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.84}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"TOKENSPEED_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.84}'
 
 sbatch_directives:
   segment: "1"
@@ -95,12 +105,12 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   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_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml

@@ -5,14 +5,14 @@ name: "kimik3-vllm-agg-b200-tp8dp2-latency-dspark-eval-agentic"
 # changes from synthetic golden AL to target-verified block.
 model:
   path: "kimik3"
-  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+    image: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
 
 dynamo:
   install: false
@@ -49,6 +49,9 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
     VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_DIRECT_DCP_A2A: "1"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
     VLLM_USE_RUST_FRONTEND: "1"
     VLLM_SERVER_DEV_MODE: "1"
     VLLM_LOG_STATS_INTERVAL: "1"
@@ -64,6 +67,8 @@ backend:
       served-model-name: "moonshotai/Kimi-K3"
       tensor-parallel-size: 8
       data-parallel-size: 2
+      decode-context-parallel-size: 8
+      dcp-comm-backend: a2a
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: fastsafetensors
@@ -73,17 +78,22 @@ backend:
       no-enable-flashinfer-autotune: true
       max-model-len: 1048576
       kv-cache-dtype: fp8
-      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      attention-backend: TOKENSPEED_MLA
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       enable-prefix-caching: true
       enable-prompt-tokens-details: true
       enable-auto-tool-choice: true
       tool-call-parser: kimi_k3
       reasoning-parser: kimi_k3
       language-model-only: true
-      max-num-seqs: 8
-      max-num-batched-tokens: 8192
+      max-num-seqs: 16
+      max-num-batched-tokens: 16384
+      prefix-match-unit: 128
+      stream-interval: 10
+      enable-cumem-allocator: true
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[8,16,24,32,40,48,56,64,72,80,88,96,104,112,120,128,256,512,1024,2048,4096,8192]}'
       # Eval must verify every draft against real target logits.
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"TOKENSPEED_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
 
 sbatch_directives:
   segment: "1"
@@ -93,12 +103,12 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   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_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-agentic.yaml

@@ -6,14 +6,14 @@ name: "kimik3-vllm-agg-b200-tp8dp2-simple-offload-dspark-agentic"
 # https://recipes.vllm.ai/moonshotai/Kimi-K3?hardware=b200&nodes=2&strategy=multi_node_dep
 model:
   path: "kimik3"
-  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+    image: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
 
 dynamo:
   install: false
@@ -50,6 +50,9 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
     VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_DIRECT_DCP_A2A: "1"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
     VLLM_USE_RUST_FRONTEND: "1"
     VLLM_SERVER_DEV_MODE: "1"
     VLLM_LOG_STATS_INTERVAL: "1"
@@ -65,6 +68,8 @@ backend:
       served-model-name: "moonshotai/Kimi-K3"
       tensor-parallel-size: 8
       data-parallel-size: 2
+      decode-context-parallel-size: 8
+      dcp-comm-backend: a2a
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: fastsafetensors
@@ -74,18 +79,23 @@ backend:
       no-enable-flashinfer-autotune: true
       max-model-len: 1048576
       kv-cache-dtype: fp8
-      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      attention-backend: TOKENSPEED_MLA
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       enable-prefix-caching: true
       enable-prompt-tokens-details: true
       enable-auto-tool-choice: true
       tool-call-parser: kimi_k3
       reasoning-parser: kimi_k3
       language-model-only: true
-      max-num-seqs: 8
-      max-num-batched-tokens: 8192
+      max-num-seqs: 96
+      max-num-batched-tokens: 16384
+      prefix-match-unit: 128
+      stream-interval: 10
+      enable-cumem-allocator: true
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136,140,144,148,152,156,160,164,168,172,176,180,184,188,192,196,200,204,208,212,216,220,224,228,232,236,240,244,248,252,256,260,264,268,272,276,280,284,288,292,296,300,304,308,312,316,320,324,328,332,336,340,344,348,352,356,360,364,368,372,376,380,384,512,1024,2048,4096,8192]}'
       # Throughput runs use the committed K7 golden synthetic AL 3.84. Eval
       # selects the paired real-verification recipe instead.
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.84}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":3,"method":"dspark","attention_backend":"TOKENSPEED_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"synthetic","synthetic_acceptance_length":3.0}'
       kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":236223201280,"lazy_offload":false}}'
 
 sbatch_directives:
@@ -96,12 +106,12 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   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_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

benchmarks/multi_node/srt-slurm-recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-vllm-simple-offload-dspark-eval-agentic.yaml

@@ -5,14 +5,14 @@ name: "kimik3-vllm-agg-b200-tp8dp2-simple-offload-dspark-eval-agentic"
 # acceptance changes from synthetic golden AL to target-verified block.
 model:
   path: "kimik3"
-  container: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+  container: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
   precision: "fp4"
 
 identity:
   model:
     repo: "moonshotai/Kimi-K3"
   container:
-    image: "vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21"
+    image: "vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf"
 
 dynamo:
   install: false
@@ -49,6 +49,9 @@ backend:
     VLLM_ENGINE_READY_TIMEOUT_S: "3600"
     VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: "1800"
     VLLM_USE_V2_MODEL_RUNNER: "1"
+    VLLM_USE_DIRECT_DCP_A2A: "1"
+    VLLM_USE_DIRECT_DCP_Q_GATHER: "1"
+    VLLM_USE_DIRECT_DCP_KV_GATHER: "1"
     VLLM_USE_RUST_FRONTEND: "1"
     VLLM_SERVER_DEV_MODE: "1"
     VLLM_LOG_STATS_INTERVAL: "1"
@@ -64,6 +67,8 @@ backend:
       served-model-name: "moonshotai/Kimi-K3"
       tensor-parallel-size: 8
       data-parallel-size: 2
+      decode-context-parallel-size: 8
+      dcp-comm-backend: a2a
       enable-expert-parallel: true
       trust-remote-code: true
       load-format: fastsafetensors
@@ -73,17 +78,22 @@ backend:
       no-enable-flashinfer-autotune: true
       max-model-len: 1048576
       kv-cache-dtype: fp8
-      attention-config: '{"use_prefill_query_quantization":true,"mla_prefill_backend":"flashinfer"}'
+      attention-backend: TOKENSPEED_MLA
+      attention-config: '{"mla_prefill_backend":"TRTLLM_RAGGED","use_prefill_query_quantization":true}'
       enable-prefix-caching: true
       enable-prompt-tokens-details: true
       enable-auto-tool-choice: true
       tool-call-parser: kimi_k3
       reasoning-parser: kimi_k3
       language-model-only: true
-      max-num-seqs: 8
-      max-num-batched-tokens: 8192
+      max-num-seqs: 96
+      max-num-batched-tokens: 16384
+      prefix-match-unit: 128
+      stream-interval: 10
+      enable-cumem-allocator: true
+      compilation-config: '{"cudagraph_mode":"FULL_AND_PIECEWISE","cudagraph_capture_sizes":[4,8,12,16,20,24,28,32,36,40,44,48,52,56,60,64,68,72,76,80,84,88,92,96,100,104,108,112,116,120,124,128,132,136,140,144,148,152,156,160,164,168,172,176,180,184,188,192,196,200,204,208,212,216,220,224,228,232,236,240,244,248,252,256,260,264,268,272,276,280,284,288,292,296,300,304,308,312,316,320,324,328,332,336,340,344,348,352,356,360,364,368,372,376,380,384,512,1024,2048,4096,8192]}'
       # Eval must verify every draft against real target logits.
-      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":7,"method":"dspark","attention_backend":"FLASHINFER_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
+      speculative-config: '{"model":"Inferact/Kimi-K3-DSpark","num_speculative_tokens":3,"method":"dspark","attention_backend":"TOKENSPEED_MLA","draft_sample_method":"probabilistic","rejection_sample_method":"block"}'
       kv-transfer-config: '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use_per_rank":236223201280,"lazy_offload":false}}'
 
 sbatch_directives:
@@ -94,12 +104,12 @@ srun_options:
 
 benchmark:
   type: custom
-  aiperf_server_metrics: true
   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_SERVER_METRICS_URLS: "http://localhost:8000/metrics"
     AIPERF_REQUIRED_SERVER_METRIC_PREFIX: "vllm:"
     RESULT_DIR: "/logs/agentic"
     PORT: "8000"

configs/nvidia-master.yaml

@@ -8039,7 +8039,7 @@ kimik3-fp4-b200-dynamo-vllm-agentic:
 # Expert parallel spans all 16 ranks, so the 896 experts remain EP16 while
 # FlashInfer fused collectives stay inside each B200 NVSwitch domain.
 kimik3-fp4-b200-dynamo-vllm-agentic-dspark:
-  image: vllm/vllm-openai:nightly-b22afe45ac797ae58e67a7a3ad79ee5714024420@sha256:144356af876edbb3a4bfee23e1444b196cc3fdadd0a0c1a7f11f721756972a21
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-5894fdf
   model: moonshotai/Kimi-K3
   model-prefix: kimik3
   runner: cluster:b200-dgxc
@@ -8051,41 +8051,9 @@ kimik3-fp4-b200-dynamo-vllm-agentic-dspark:
     agentic-coding:
     - dram-utilization: 0.63
       search-space:
-      # Low-latency and high-interactivity points.
-      - spec-decoding: mtp
-        conc-list: [1, 2, 4]
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 16
-          dp-attn: true
-          additional-settings:
-          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml"
-          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml"
-        decode:
-          num-worker: 0
-          tp: 8
-          ep: 16
-          dp-attn: true
-      # Balanced medium-concurrency points.
+      # Retained resident latency curve from the completed broad fast sweep.
       - spec-decoding: mtp
-        conc-list: [8, 12, 16]
-        prefill:
-          num-worker: 1
-          tp: 8
-          ep: 16
-          dp-attn: true
-          additional-settings:
-          - "CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-agentic.yaml"
-          - "EVAL_CONFIG_FILE=recipes/vllm/kimi-k3/agentic/agg-b200-tp8dp2-latency-dspark-eval-agentic.yaml"
-        decode:
-          num-worker: 0
-          tp: 8
-          ep: 16
-          dp-attn: true
-      # GPU-resident throughput points around the prior c16-c32 KV cliff.
-      - spec-decoding: mtp
-        conc-list: [20, 24, 28, 32]
+        conc-list: [1, 2, 4, 8]
         prefill:
           num-worker: 1
           tp: 8
@@ -8099,12 +8067,12 @@ kimik3-fp4-b200-dynamo-vllm-agentic-dspark:
           tp: 8
           ep: 16
           dp-attn: true
-      # CPU KV-offload crossover and capacity points. Keep the resident points
-      # above so the same concurrency can be compared with one variable changed.
+      # Retain the measured SimpleCPUOffloadConnector knee, immediate boundary,
+      # and high-capacity controls from the completed broad fast sweep.
       - spec-decoding: mtp
         kv-offloading: dram
         kv-offload-backend: { name: vllm-simple, version: "13c59a3" }
-        conc-list: [8, 12, 16, 20, 24, 28, 32, 48, 64]
+        conc-list: [16, 24, 32, 40, 48]
         prefill:
           num-worker: 1
           tp: 8

perf-changelog.yaml

@@ -5818,3 +5818,20 @@
   description:
     - "Extend the SimpleCPUOffloadConnector grid to c8/c12/c16/c20/c24/c28/c32/c48/c64 to locate its crossover against the resident curve"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2475
+
+- config-keys:
+    - kimik3-fp4-b200-dynamo-vllm-agentic-dspark
+  description:
+    - "Fix AIPerf server-metrics configuration for the pinned Kimi K3 srt-slurm renderer by passing the aggregate vLLM endpoint through the supported custom-benchmark environment contract"
+    - "After the complete 19-point AgentX-fast search, retain resident c2/c4/c8/c12 and SimpleCPUOffloadConnector c8/c12/c16/c28/c32/c48 for the full-duration latency curve, knee, cliff boundary, and high-capacity controls"
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2569
+
+- config-keys:
+    - kimik3-fp4-b200-dynamo-vllm-agentic-dspark
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Add decode context parallelism 8 within each TP8 replica, with the a2a backend and the direct DCP kernels."
+    - "Serve on TokenspeedMLA with TRTLLM_RAGGED prefill, a 16384 prefill token budget, prefix-match-unit 128 and an explicit FULL_AND_PIECEWISE capture list sized to each arm."
+    - "Draft with DSpark level 7 at concurrency 8 and below and level 3 above it, each at its committed golden acceptance length, and move to the image carrying the Kimi K3 DCP, DSpark-under-DCP and Mooncake changes."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2618