[codex] Update MiniMax M3 B200 EAGLE3 FlashInfer image

.github/configs/nvidia-master.yaml

@@ -11959,13 +11959,12 @@ minimaxm3-fp8-b200-vllm:
 # EAGLE3 speculative-decoding (spec-decoding: mtp) variant of
 # minimaxm3-fp8-b200-vllm, pairing MiniMaxAI/MiniMax-M3-MXFP8 with the
 # Inferact/MiniMax-M3-EAGLE3 draft head (3 speculative tokens, drafter pinned
-# to FLASH_ATTN since the head is MHA and FlashInfer needs GQA/MQA at page
-# size 128). Search space mirrors the non-MTP entry trimmed at the
+# to TRITON_ATTN). Search space mirrors the non-MTP entry trimmed at the
 # extreme-concurrency end, identical to the minimaxm3-fp8-b300-vllm-mtp
 # precedent: spec decode pays off at low/mid concurrency while acceptance
 # dilutes in big batches, and the draft weights + draft KV shave headroom.
 minimaxm3-fp8-b200-vllm-mtp:
-  image: vllm/vllm-openai:minimax-m3
+  image: vllm/vllm-openai:minimax-m3-0618-x86_64-cu130
   model: MiniMaxAI/MiniMax-M3-MXFP8
   model-prefix: minimaxm3
   runner: b200-dgxc

benchmarks/single_node/fixed_seq_len/minimaxm3_fp8_b200_mtp.sh

@@ -10,12 +10,8 @@
 # encoder for the text-only benchmark. dp-attn=true maps to DP×EP (DEP);
 # ep>1 maps to TP+EP (TEP).
 #
-# The drafter is pinned to FLASH_ATTN: the EAGLE3 head is MHA, and FlashInfer
-# only supports page size 128 through its trtllm-gen kernel, which requires
-# GQA/MQA — engine init dies in FlashInferMetadataBuilder otherwise (the
-# failure hit on the B300 MTP canary). The target keeps its default
-# (FlashInfer) backend; FLASH_ATTN takes any multiple-of-16 block size, so
-# the mandatory 128 is fine for the draft.
+# The target uses the FlashInfer TRT-LLM attention path. The EAGLE3 drafter is
+# pinned separately to TRITON_ATTN.
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -31,6 +27,39 @@ check_env_vars \
     RANDOM_RANGE_RATIO \
     RESULT_FILENAME
 
+# The 0618 image keeps MiniMax M3 top-k indices in a persistent
+# [head_kv, max_tokens, topK] buffer for CUDA graphs. Slicing that buffer to
+# the actual prefill length is non-contiguous when TP leaves multiple local KV
+# heads, and the MSA CSR builder rejects it. Materialize the slice until the
+# image includes this fix.
+python3 - <<'PYEOF' || { echo "MiniMax M3 MSA contiguity patch failed" >&2; exit 1; }
+import importlib.util
+import pathlib
+
+spec = importlib.util.find_spec("vllm")
+if spec is None or not spec.submodule_search_locations:
+    raise RuntimeError("Could not locate the installed vllm package")
+
+target = (
+    pathlib.Path(next(iter(spec.submodule_search_locations)))
+    / "models"
+    / "minimax_m3"
+    / "nvidia"
+    / "sparse_attention_msa.py"
+)
+src = target.read_text()
+old = "            prefill_topk = topk[:, nd:num_tokens, :]\n"
+new = "            prefill_topk = topk[:, nd:num_tokens, :].contiguous()\n"
+
+if new in src:
+    print(f"[minimax-m3-msa-patch] already applied: {target}")
+elif src.count(old) == 1:
+    target.write_text(src.replace(old, new, 1))
+    print(f"[minimax-m3-msa-patch] patched: {target}")
+else:
+    raise RuntimeError(f"Expected exactly one patch anchor in {target}")
+PYEOF
+
 DRAFT_MODEL="Inferact/MiniMax-M3-EAGLE3"
 
 if [[ -n "$SLURM_JOB_ID" ]]; then
@@ -68,7 +97,7 @@ if [ "${DP_ATTENTION}" = "true" ]; then
 elif [ "$EP_SIZE" -gt 1 ]; then
   PARALLEL_ARGS="--tensor-parallel-size=$TP --enable-expert-parallel"
 else
-  PARALLEL_ARGS="--tensor-parallel-size=$TP --moe-backend marlin"
+  PARALLEL_ARGS="--tensor-parallel-size=$TP"
 fi
 
 # use 3 speculative tokens for all configs for now
@@ -87,10 +116,13 @@ $PARALLEL_ARGS \
 --gpu-memory-utilization 0.90 \
 --max-model-len $MAX_MODEL_LEN \
 --block-size 128 \
+--attention-config '{"backend": "FLASHINFER", "use_trtllm_attention": true}' \
+--attention-config.indexer_kv_dtype "fp8" \
+--kv-cache-dtype fp8 \
 --language-model-only \
 --max-cudagraph-capture-size 2048 \
 --max-num-batched-tokens "$((ISL * 2 ))" \
---speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"FLASH_ATTN\"}" \
+--speculative-config "{\"method\": \"eagle3\", \"model\": \"$DRAFT_MODEL_PATH\", \"num_speculative_tokens\": $NUM_SPEC_TOKENS, \"attention_backend\": \"TRITON_ATTN\"}" \
 --stream-interval 20 --no-enable-prefix-caching \
 --trust-remote-code > $SERVER_LOG 2>&1 &
 

perf-changelog.yaml

@@ -3951,6 +3951,15 @@
     - "Update Applied TBO on high concurrencies"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1717
 
+- config-keys:
+    - minimaxm3-fp8-b200-vllm-mtp
+  description:
+    - "Update the MiniMax-M3 B200 single-node image to vllm/vllm-openai:minimax-m3-0618-x86_64-cu130."
+    - "Enable FlashInfer TRT-LLM attention with FP8 indexer KV and KV cache; use TRITON_ATTN for the EAGLE3 drafter."
+    - "Switch TP-only configurations from explicit Marlin MoE to the new image's default FlashInfer TRT-LLM MoE backend."
+    - "Patch the image's MiniMax M3 MSA prefill path to materialize sliced top-k indices before CSR construction."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/1832
+
 - config-keys:
     - minimaxm3-fp8-b200-vllm
   description: