Tune MiniMax-M3 AgentX on B200 and B300 with updated vLLM / 更新 B200 与 B300 上的 MiniMax-M3 AgentX 配置

benchmarks/single_node/agentic/minimaxm3_fp4_b200_mtp.sh

@@ -11,10 +11,11 @@ set -x
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
-check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION PORT EVAL_ONLY
 
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 NUM_SPEC_TOKENS=3
+TOKENS_PER_SEQ=$((1 + NUM_SPEC_TOKENS))
 # Golden AL for the GQA draft head: golden_al_distribution/minimaxm3_eagle3_gqa.yaml
 # minimax-m3.thinking_on[3]. The non-GQA curve (minimaxm3_eagle3.yaml) reads 2.83.
 SYNTHETIC_ACCEPT_LEN=2.78
@@ -103,8 +104,8 @@ install_agentic_deps
 
 OFFLOAD_ARGS=()
 if require_agentic_kv_offload_backend vllm-simple; then
-    python3 "$(dirname "$0")/../../../runners/patch_vllm_simple_kv_offload.py"
-    CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1024 * 1024 * 1024))
+    # The matrix emits decimal GB; SimpleCPUOffloadConnector expects bytes.
+    CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000))
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_CONFIG=$(printf \
         '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":%d,"lazy_offload":true}}' \
@@ -116,10 +117,18 @@ export PYTHONNOUSERSITE=1
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export VLLM_FLOAT32_MATMUL_PRECISION=high
 export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
-
-# B200's 180 GB leaves little beyond the ~250 GB checkpoint: TP2 cannot
-# host 1M-context KV for one request at 0.9, so TP4 is the smallest topology.
-GPU_MEMORY_UTILIZATION="0.9"
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
+
+# Reserve 5% of HBM for runtime spikes while giving the KV pool more room.
+# TP2 failed the full-context startup check at 0.90; it needs a separate GPU
+# qualification before joining the sweep at this higher utilization.
+GPU_MEMORY_UTILIZATION="0.95"
+MAX_NUM_SEQS=$((2 * CONC))
+MAX_CUDAGRAPH_CAPTURE_SIZE=$((TOKENS_PER_SEQ * MAX_NUM_SEQS))
+if (( MAX_CUDAGRAPH_CAPTURE_SIZE > 512 )); then
+    MAX_CUDAGRAPH_CAPTURE_SIZE=512
+fi
 
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
@@ -138,11 +147,11 @@ trap 'exit 143' TERM
 
 if [ "${EVAL_ONLY:-}" = "true" ]; then
     SPEC_CONFIG=$(printf \
-        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN"}' \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"TRITON_ATTN","kv_cache_dtype":"fp8"}' \
         "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS")
 else
     SPEC_CONFIG=$(printf \
-        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"TRITON_ATTN","kv_cache_dtype":"fp8","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
         "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS" "$SYNTHETIC_ACCEPT_LEN")
 fi
 
@@ -157,14 +166,16 @@ VLLM_CMD=(
     --block-size 128
     --language-model-only
     --enable-prefix-caching
+    --enable-chunked-prefill
     --no-enable-flashinfer-autotune
     --reasoning-parser minimax_m3
     --tool-call-parser minimax_m3
     --enable-auto-tool-choice
     --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
     --attention-config '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}'
     --kv-cache-dtype fp8
-    --max-cudagraph-capture-size 512
+    --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE"
+    --max-num-seqs "$MAX_NUM_SEQS"
     --max-num-batched-tokens 16384
     --stream-interval 20
     --trust-remote-code

benchmarks/single_node/agentic/minimaxm3_fp4_b300_mtp.sh

@@ -7,10 +7,11 @@ set -x
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
-check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION
+check_env_vars MODEL TP CONC KV_OFFLOADING TOTAL_CPU_DRAM_GB RESULT_DIR DURATION PORT EVAL_ONLY
 
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3-GQA"
 NUM_SPEC_TOKENS=3
+TOKENS_PER_SEQ=$((1 + NUM_SPEC_TOKENS))
 # Golden AL for the GQA draft head: golden_al_distribution/minimaxm3_eagle3_gqa.yaml
 # minimax-m3.thinking_on[3]. The non-GQA curve (minimaxm3_eagle3.yaml) reads 2.83.
 SYNTHETIC_ACCEPT_LEN=2.78
@@ -40,8 +41,8 @@ install_agentic_deps
 
 OFFLOAD_ARGS=()
 if require_agentic_kv_offload_backend vllm-simple; then
-    python3 "$(dirname "$0")/../../../runners/patch_vllm_simple_kv_offload.py"
-    CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1024 * 1024 * 1024))
+    # The matrix emits decimal GB; SimpleCPUOffloadConnector expects bytes.
+    CPU_OFFLOAD_BYTES=$((TOTAL_CPU_DRAM_GB * 1000 * 1000 * 1000))
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_CONFIG=$(printf \
         '{"kv_connector":"SimpleCPUOffloadConnector","kv_role":"kv_both","kv_connector_extra_config":{"cpu_bytes_to_use":%d,"lazy_offload":true}}' \
@@ -53,6 +54,14 @@ export PYTHONNOUSERSITE=1
 export VLLM_ENGINE_READY_TIMEOUT_S=3600
 export VLLM_FLOAT32_MATMUL_PRECISION=high
 export VLLM_FLASHINFER_ALLREDUCE_BACKEND=trtllm
+export AIPERF_SERVER_METRICS_URLS="http://localhost:${PORT}/metrics"
+export AIPERF_REQUIRED_SERVER_METRIC_PREFIX="vllm:"
+
+MAX_NUM_SEQS=$((2 * CONC))
+MAX_CUDAGRAPH_CAPTURE_SIZE=$((TOKENS_PER_SEQ * MAX_NUM_SEQS))
+if (( MAX_CUDAGRAPH_CAPTURE_SIZE > 512 )); then
+    MAX_CUDAGRAPH_CAPTURE_SIZE=512
+fi
 
 SERVER_LOG="$RESULT_DIR/server.log"
 mkdir -p "$RESULT_DIR"
@@ -71,11 +80,11 @@ trap 'exit 143' TERM
 
 if [ "${EVAL_ONLY:-}" = "true" ]; then
     SPEC_CONFIG=$(printf \
-        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN"}' \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"TRITON_ATTN","kv_cache_dtype":"fp8"}' \
         "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS")
 else
     SPEC_CONFIG=$(printf \
-        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"FLASH_ATTN","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
+        '{"method":"eagle3","model":"%s","num_speculative_tokens":%d,"attention_backend":"TRITON_ATTN","kv_cache_dtype":"fp8","rejection_sample_method":"synthetic","synthetic_acceptance_length":%.2f}' \
         "$DRAFT_MODEL_PATH" "$NUM_SPEC_TOKENS" "$SYNTHETIC_ACCEPT_LEN")
 fi
 
@@ -86,19 +95,21 @@ VLLM_CMD=(
     --host 0.0.0.0
     --port "$PORT"
     --tensor-parallel-size "$TP"
-    --gpu-memory-utilization 0.9
+    --gpu-memory-utilization 0.95
     --block-size 128
     --language-model-only
     --enable-prefix-caching
+    --enable-chunked-prefill
     --no-enable-flashinfer-autotune
     --reasoning-parser minimax_m3
     --tool-call-parser minimax_m3
     --enable-auto-tool-choice
     --default-chat-template-kwargs '{"thinking_mode":"enabled"}'
     --attention-config '{"backend":"FLASHINFER","use_trtllm_attention":true,"indexer_kv_dtype":"fp8"}'
     --kv-cache-dtype fp8
-    --max-cudagraph-capture-size 512
-    --max-num-batched-tokens 16384
+    --max-cudagraph-capture-size "$MAX_CUDAGRAPH_CAPTURE_SIZE"
+    --max-num-seqs "$MAX_NUM_SEQS"
+    --max-num-batched-tokens 32768
     --stream-interval 20
     --trust-remote-code
     --speculative-config "$SPEC_CONFIG"

configs/nvidia-master.yaml

@@ -5562,7 +5562,7 @@ qwen3.5-fp4-gb200-dynamo-sglang-agentic-mtp:
           additional-settings:
           - "CONFIG_FILE=recipes/qwen3.5/sglang/gb200-fp4/agentx/agg-tp2ep2-mtp-hicache-cap48.yaml"
 minimaxm3-fp4-b300-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45
+  image: vllm/vllm-openai:nightly-29468dde8b515031dc6d4d9d06bf0a2fa0442098
   model: nvidia/MiniMax-M3-NVFP4
   model-prefix: minimaxm3
   runner: cluster:b300-dsxe
@@ -5571,14 +5571,14 @@ minimaxm3-fp4-b300-vllm-agentic-mtp:
   multinode: false
   scenarios:
     agentic-coding:
-    - dram-utilization: 0.683
+    - dram-utilization: 1.0
       search-space:
       - { tp: 8, spec-decoding: mtp, kv-offloading: none, conc-list: [1] }
       - { tp: 4, spec-decoding: mtp, kv-offloading: none, conc-list: [1, 5, 10, 15, 20] }
-      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [30] }
+      - { tp: 4, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [30, 40, 48, 64] }
     - dram-utilization: 1.0
       search-space:
-      - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [24] }
+      - { tp: 2, spec-decoding: mtp, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [24, 32, 40] }
 
 minimaxm3-fp4-gb300-dynamo-vllm-agentic-mtp-disagg:
   image: vllm/vllm-openai:nightly-5e35a6f4f9bbc217c599692157ca985c894373f7
@@ -5690,33 +5690,34 @@ minimaxm3-fp4-b300-trtllm-agentic-mtp:
 # MiniMax-M3's only active B200 configuration -- the 8k1k scenario that carried
 # the previous B200 curves was removed in #2493.
 #
-# TP4-only. The B300 sibling also runs a TP2 arm, but TP2 does not fit on B200:
+# TP4-only until a new TP2 arm is qualified. At 0.90 HBM utilization TP2
+# does not fit on B200:
 # the ~250 GB NVFP4 checkpoint leaves a 31.72 GiB KV pool at
 # --gpu-memory-utilization 0.9, and one request at the model's 1,048,576-token
 # max_model_len needs 37.62 GiB, so vLLM raises in _check_enough_kv_cache_memory
 # and the engine never reaches serving (run 31141913741 -- both TP2 cells, c1
-# and c2, died identically at init; every TP4 cell passed). This is structural,
-# not a concurrency cliff: the estimated max model length at TP2 is 883,840,
-# below the corpus context, so no conc-list would have made the arm run.
+# and c2, died identically at init; every TP4 cell passed). This is a startup
+# capacity limit at 0.90, not a concurrency cliff: the estimated max model
+# length at TP2 is 883,840, below the corpus context. The new 0.95 limit may
+# clear that check, but needs a live full-context run.
 #
-# Concurrency on the surviving arms still differs from B300, whose TP4 pool is
-# ~787 GB against B200's ~398 GB:
+# Concurrency on the surviving arms still differs from B300; at the prior
+# 0.90 HBM limit, its TP4 KV pool was ~787 GB against B200's ~398 GB:
 #   TP4   B300 runs 1-20; B200 samples 5-15 densely because the cliff lands
 #         inside that range at roughly half the KV, and retains 20 past it.
 #   TP4 + SimpleCPU  picks up at 20 rather than B300's 30, since the host tier
 #         has to start absorbing the working set one step earlier, and stops at
-#         40 rather than following B300 to 75. Past 40 the host tier is
+#         40 rather than following B300's new sweep to 64. Past 40 the host tier is
 #         absorbing a working set that never fit in B200's ~398 GB TP4 pool to
 #         begin with, so those points cost four engine starts each to trace a
 #         tail that is offload bandwidth, not the SKU.
 #
-# dram-utilization 0.683 is the B300 value verbatim, and it resolves to the same
-# 1,024 GB engine-level CPU KV budget at TP4: b200-nscale and b300-nv both report
-# more installed DRAM than MAX_AGENTIC_AVAILABLE_CPU_DRAM_MIB, so both clamp to
-# the same 3 TB AgentX ceiling before the proportional-GPU rule is applied.
-# GPU-resident points receive a zero budget.
+# B200's 1.0 offload arm uses the full declared host budget: 2,063,920 MiB
+# times TP4/8, or about 1,082 GB for this server. GPU-resident points receive
+# no host KV budget. B300 has a higher, 3 TB capped node budget (1,499 GB for
+# TP4, 749 GB for TP2 after whole-GB rounding).
 minimaxm3-fp4-b200-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45
+  image: vllm/vllm-openai:nightly-29468dde8b515031dc6d4d9d06bf0a2fa0442098
   model: nvidia/MiniMax-M3-NVFP4
   model-prefix: minimaxm3
   runner: cluster:b200-nscale

perf-changelog.yaml

@@ -8950,3 +8950,40 @@
   description:
     - "Update SGLang image from v0.5.19-cu130 to v0.5.20-cu130 and pass --cuda-graph-max-bs-decode after the alias removal."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3334
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-agentic-mtp
+    - minimaxm3-fp4-b300-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Pin MiniMax-M3 B200/B300 AgentX to vLLM nightly-29468dde8b515031dc6d4d9d06bf0a2fa0442098 and remove the obsolete SimpleCPU KV patch."
+    - "Increase HBM and B300 host KV budgets, tune serving limits, expand the B300 sweep, and correct host-capacity units."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3435
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-agentic-mtp
+    - minimaxm3-fp4-b300-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use BF16 EAGLE3 draft KV cache to avoid the FA4 FP8 descale startup failure."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3435
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-agentic-mtp
+    - minimaxm3-fp4-b300-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use FlashInfer with FP8 EAGLE3 draft KV cache on B200 and B300."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3435
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-agentic-mtp
+    - minimaxm3-fp4-b300-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Use Triton attention with FP8 EAGLE3 draft KV cache."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3435