Retune DSV4 B300 AgentX MTP sweep / 调优 DSV4 B300 AgentX MTP 扫描

benchmarks/single_node/agentic/dsv4_fp4_b300_vllm_mtp.sh

@@ -18,8 +18,8 @@ set -x
 # Required env vars:
 #   MODEL, TP, CONC, KV_OFFLOADING, TOTAL_CPU_DRAM_GB, RESULT_DIR
 #
-# TP4, TP8, and DEP8 (TP8 + DP-attention) are GPU-resident (KV_OFFLOADING=none).
-# DEP4 uses KV_OFFLOADING=dram with KV_OFFLOAD_BACKEND=vllm-simple or mooncake.
+# TP8 and TP4 c8 are GPU-resident. TP4 c16, DEP4, and DEP8 use DRAM offload
+# with KV_OFFLOAD_BACKEND=vllm-simple or mooncake.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -49,9 +49,9 @@ if [ "$DP_ATTENTION" = "true" ] && [ $((2 * CONC % TP)) -ne 0 ]; then
     exit 1
 fi
 
-# DEP8 (TP8 + DP-attention) is a GPU-resident, high-concurrency arm that is
-# tuned separately from the smaller DEP4 arm (larger prefill token budget,
-# long-prefill chunking, and a lower GPU-memory-utilization headroom).
+# DEP8 (TP8 + DP-attention) is a high-concurrency SimpleCPU arm tuned separately
+# from DEP4 with a larger prefill token budget and lower GPU-memory-utilization
+# headroom. Both DEP arms chunk long prefills.
 IS_DEP8=false
 if [ "$DP_ATTENTION" = "true" ] && [ "$TP" -eq 8 ]; then
     IS_DEP8=true
@@ -119,8 +119,8 @@ ROUTER_PID=""
 MOONCAKE_MASTER_PID=""
 
 # The generated TOTAL_CPU_DRAM_GB budget is proportional to allocated GPUs.
-# On cluster:b300-nv, dram-utilization=0.80 and DEP4 resolve to roughly the
-# source recipe's 280 GiB per DP rank. TP4 remains GPU-resident.
+# On cluster:b300-nv, dram-utilization=0.95 gives both DEP4 and DEP8 356 GB per
+# DP rank (1,424 GB and 2,849 GB total, respectively). TP arms remain GPU-resident.
 OFFLOAD_ARGS=()
 case "$KV_OFFLOAD_BACKEND" in
     "")
@@ -238,14 +238,14 @@ if [ "$EP_SIZE" -gt 1 ]; then
     )
 fi
 if [ "$DP_ATTENTION" = "true" ]; then
-    MODE_ARGS+=(--prefill-schedule-interval 8)
+    MODE_ARGS+=(
+        --prefill-schedule-interval 8
+        --long-prefill-token-threshold 512
+    )
     if [ "$IS_DEP8" = "true" ]; then
-        # GPU-resident DEP8 gets a larger prefill token budget and chunks long
-        # prefills so decode latency stays bounded at high concurrency.
-        MODE_ARGS+=(
-            --max-num-batched-tokens 16384
-            --long-prefill-token-threshold 4096
-        )
+        # DEP8 gets a larger prefill token budget; the shared long-prefill
+        # threshold keeps decode latency bounded under load.
+        MODE_ARGS+=(--max-num-batched-tokens 16384)
     else
         MODE_ARGS+=(--max-num-batched-tokens 8192)
     fi

configs/nvidia-master.yaml

@@ -1725,7 +1725,7 @@ dsv4-fp4-b300-vllm:
       - { tp: 8, ep: 8, dp-attn: true, conc-start: 2048, conc-end: 2048 }
 
 dsv4-fp4-b300-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-904e4ec
+  image: vllm/vllm-openai:nightly-dev-x86_64-cu13.0.1-426e59f
   model: deepseek-ai/DeepSeek-V4-Pro
   model-prefix: dsv4
   runner: cluster:b300-nv
@@ -1734,18 +1734,18 @@ dsv4-fp4-b300-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.80
+    - dram-utilization: 0.95
       search-space:
       # TP8 GPU-resident + MTP (num_speculative_tokens=3)
-      - { tp: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8] }
+      - { tp: 8, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 4] }
       # TP4 GPU-resident + MTP (num_speculative_tokens=3)
-      - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [1, 2, 4, 6, 8, 12, 16, 20] }
+      - { tp: 4, kv-offloading: none, spec-decoding: mtp, conc-list: [8] }
       # TP4 SimpleCPU + MTP (num_speculative_tokens=3)
-      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [20, 24, 28, 32, 36, 40] }
+      - { tp: 4, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [16] }
       # DEP4 SimpleCPU + MTP (num_speculative_tokens=3)
-      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [32, 40, 48, 56], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 4, ep: 4, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [48, 64], router: { name: vllm-router, version: "0.1.14" } }
       # DEP8 SimpleCPU + MTP (num_speculative_tokens=3)
-      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: none, spec-decoding: mtp, conc-list: [64, 96, 112, 128, 144, 160, 176, 192, 224], router: { name: vllm-router, version: "0.1.14" } }
+      - { tp: 8, ep: 8, dp-attn: true, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [128, 256, 384, 512, 576], router: { name: vllm-router, version: "0.1.14" } }
 
 dsv4-fp4-b300-trt:
   image: ghcr.io#semianalysisai/trtllm-deepseek-v4:feat-deepseek_v4-c185066

perf-changelog.yaml

@@ -5770,3 +5770,10 @@
     - "models.yaml: add mtp_flags (--speculative-algorithm EAGLE --speculative-eagle-topk 1) to DeepSeek-V4-Pro-AgentX; num-steps/num-draft-tokens are derived from DECODE_MTP_SIZE in server_sglang.sh. DeepSeek-V4's MTP head is a native NextN/EAGLE draft shipped with the model, so no --speculative-draft-model-path is needed."
     - "Two search-space arms: TP8/EP1 no-DP at conc-list [2,4,8,16,32], and TP8/EP8/DPA (ep=8, dp-attn=true) at conc-list [64,96,128], both on image lmsysorg/sglang-rocm:v0.5.15.post1-rocm720-mi35x-20260719."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2309
+
+- config-keys:
+    - dsv4-fp4-b300-vllm-agentic-mtp
+  description:
+    - "Update the B300 DSV4 AgentX MTP image and SimpleCPU sweep through c576."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2536
+