chore: adjust conc list

Author: ivanium

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep12-agentic.yaml

@@ -134,7 +134,7 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 120
+      max-num-seqs: 256
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-1p1d-dep8-dep8-agentic.yaml

@@ -134,7 +134,7 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 120
+      max-num-seqs: 128
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
@@ -158,13 +158,13 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 60
+      max-num-seqs: 64
       max-num-batched-tokens: 256
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
       compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
-      max-cudagraph-capture-size: 60
+      max-cudagraph-capture-size: 64
       gpu-memory-utilization: 0.95
       no-disable-hybrid-kv-cache-manager: true
       tokenizer-mode: deepseek_v4

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-2p1d-dep8-dep12-agentic.yaml

@@ -134,7 +134,7 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 120
+      max-num-seqs: 192
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true

benchmarks/multi_node/srt-slurm-recipes/vllm/deepseek-v4/agentic/disagg-gb200-3p1d-dep8-dep16-agentic.yaml

@@ -134,7 +134,7 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 120
+      max-num-seqs: 214
       max-num-batched-tokens: 8192
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
@@ -158,13 +158,13 @@ backend:
       numa-bind: true
       attention-config: '{"backend": "FLASHINFER_MLA_SPARSE_DSV4", "use_prefill_query_quantization": true, "use_fp4_indexer_cache": true}'
       max-model-len: 1048576
-      max-num-seqs: 128
+      max-num-seqs: 160
       max-num-batched-tokens: 256
       trust-remote-code: true
       no-enable-flashinfer-autotune: true
       block-size: 256
       compilation-config: '{"cudagraph_mode":"FULL_DECODE_ONLY","mode":0}'
-      max-cudagraph-capture-size: 128
+      max-cudagraph-capture-size: 160
       gpu-memory-utilization: 0.95
       no-disable-hybrid-kv-cache-manager: true
       tokenizer-mode: deepseek_v4

configs/nvidia-master.yaml

@@ -7708,7 +7708,7 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-1p1d-dep8-dep8:
     agentic-coding:
     - search-space:
       - spec-decoding: none
-        conc-list: [64, 128, 256, 320]
+        conc-list: [64, 128, 192, 256]
         prefill:
           num-worker: 1
           tp: 8
@@ -7737,7 +7737,7 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-1p1d-dep8-dep12:
     agentic-coding:
     - search-space:
       - spec-decoding: none
-        conc-list: [128, 256, 320]
+        conc-list: [384, 512]
         prefill:
           num-worker: 1
           tp: 8
@@ -7766,7 +7766,7 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-2p1d-dep8-dep12:
     agentic-coding:
     - search-space:
       - spec-decoding: none
-        conc-list: [480, 640, 768]
+        conc-list: [640, 720, 768]
         prefill:
           num-worker: 2
           tp: 8
@@ -7795,7 +7795,7 @@ dsv4-fp4-gb200-dynamo-vllm-agentic-3p1d-dep8-dep16:
     agentic-coding:
     - search-space:
       - spec-decoding: none
-        conc-list: [640, 800, 960, 1280]
+        conc-list: [800, 960, 1024, 1280]
         prefill:
           num-worker: 3
           tp: 8