feat(amd): add Kimi-K3 DSpark AgentX variant

Add the MI355X DSpark wrapper, TP8 vLLM configuration, synthetic golden-AL handling, and performance changelog entry.

中文：新增 MI355X Kimi-K3 DSpark AgentX 包装脚本与 TP8 vLLM 配置，接入合成黄金接受长度，并登记性能变更日志。

Author: edwingao28

.github/workflows/e2e-tests.yml

@@ -310,14 +310,15 @@ jobs:
             dp-attn: ${{ matrix.config.dp-attn }}
             conc: ${{ matrix.config.conc }}
             kv-offloading: ${{ matrix.config.kv-offloading }}
-            kv-offload-backend: ${{ matrix.config.kv-offload-backend }}
+            kv-offload-backend: ${{ matrix.config['kv-offload-backend'].name }}
+            kv-offload-backend-metadata: ${{ matrix.config['kv-offload-backend'] && toJson(matrix.config['kv-offload-backend']) || '' }}
             total-cpu-dram-gb: ${{ matrix.config.total-cpu-dram-gb }}
             duration: ${{ inputs.agentx-fast && '1200' || (inputs.duration-override != '' && inputs.duration-override || matrix.config.duration) }}
             agentx-fast: ${{ inputs.agentx-fast }}
             isl: '0'
             osl: '0'
             max-model-len: '0'
-            spec-decoding: 'none'
+            spec-decoding: ${{ matrix.config.spec-decoding }}
             disagg: ${{ 'false' }}
             run-eval: true
             eval-only: true

benchmarks/single_node/agentic/kimik3_fp4_mi355x.sh

@@ -47,7 +47,8 @@ set -x
 #   LANGUAGE_MODEL_ONLY      false  (reference loads the vision tower)
 #   KV_CACHE_DTYPE           fp8    (default for every arm; =auto for a bf16 A/B)
 #   MAX_MODEL_LEN            unset  (unset -> vLLM derives K3's 1M context)
-#   SPEC_DECODE              false  (DSpark; UNVALIDATED on ROCm)
+#   SPEC_DECODE              false  (enabled by the _mtp DSpark wrapper)
+#   ENFORCE_EAGER            false  (enabled by the _mtp DSpark wrapper)
 
 source "$(dirname "$0")/../../benchmark_lib.sh"
 
@@ -543,6 +544,11 @@ if [ -n "${MAX_MODEL_LEN:-}" ] && [ "${MAX_MODEL_LEN}" != "0" ]; then
     MAX_MODEL_LEN_ARGS=(--max-model-len "$MAX_MODEL_LEN")
 fi
 
+EAGER_ARGS=()
+if [ "${ENFORCE_EAGER:-false}" = "true" ]; then
+    EAGER_ARGS=(--enforce-eager)
+fi
+
 # The reference command passes neither --enable-prefix-caching nor
 # --no-enable-prefix-caching, and this build's default is None (vLLM decides
 # internally), so by default we pass nothing and stay aligned. Two reasons this
@@ -579,17 +585,36 @@ fi
 
 # The upstream DSpark config pins "attention_backend": "FLASHINFER_MLA", which
 # is CUDA-only and cannot be used verbatim on gfx950; SPEC_ATTN_BACKEND
-# overrides it. Golden AL on B300 is 3.78 at 7 draft tokens
-# (golden_al_distribution/kimik3_dspark.yaml), so this is the largest decode-side
-# lever if it can be made to run here.
+# overrides it. AgentX throughput uses synthetic acceptance pinned to the
+# committed golden curve. Eval-only runs use real block verification because
+# synthetic acceptance bypasses correctness verification.
 SPEC_ARGS=()
 if [ "${SPEC_DECODE:-false}" = "true" ]; then
     SPEC_DRAFT_MODEL="${SPEC_DRAFT_MODEL:-Inferact/Kimi-K3-DSpark}"
     SPEC_NUM_TOKENS="${SPEC_NUM_TOKENS:-7}"
     SPEC_ATTN_BACKEND="${SPEC_ATTN_BACKEND:-TRITON_MLA}"
+    SPEC_GOLDEN_AL="${SPEC_GOLDEN_AL:-3.84}"
+    SPEC_GOLDEN_AL_FILE="${SPEC_GOLDEN_AL_FILE:-golden_al_distribution/kimik3_dspark_probabilistic_sample_method_block_rejection_sample_method.yaml}"
+    FILE_GOLDEN_AL=$(awk -v k="$SPEC_NUM_TOKENS" '
+        /^kimi-k3:/                { in_model = 1; next }
+        /^[^[:space:]#]/           { in_model = 0 }
+        in_model && /thinking_on:/ { in_thinking = 1; next }
+        in_thinking && $1 == k":"  { print $2; exit }
+    ' "$SPEC_GOLDEN_AL_FILE" 2>/dev/null)
+    if [ "$FILE_GOLDEN_AL" != "$SPEC_GOLDEN_AL" ]; then
+        echo "Error: DSpark golden AL mismatch: $SPEC_GOLDEN_AL_FILE gives" >&2
+        echo "       '${FILE_GOLDEN_AL:-<unreadable>}' for k=$SPEC_NUM_TOKENS," >&2
+        echo "       but the recipe pins $SPEC_GOLDEN_AL." >&2
+        exit 1
+    fi
+    if [ "${EVAL_ONLY:-false}" = "true" ]; then
+        SPEC_REJECTION_CONFIG="\"rejection_sample_method\":\"block\""
+    else
+        SPEC_REJECTION_CONFIG="\"rejection_sample_method\":\"synthetic\",\"synthetic_acceptance_length\":$SPEC_GOLDEN_AL"
+    fi
     SPEC_ARGS=(
         --speculative-config
-        "{\"model\":\"$SPEC_DRAFT_MODEL\",\"num_speculative_tokens\":$SPEC_NUM_TOKENS,\"method\":\"dspark\",\"attention_backend\":\"$SPEC_ATTN_BACKEND\",\"draft_sample_method\":\"probabilistic\",\"rejection_sample_method\":\"block\"}"
+        "{\"model\":\"$SPEC_DRAFT_MODEL\",\"num_speculative_tokens\":$SPEC_NUM_TOKENS,\"method\":\"dspark\",\"attention_backend\":\"$SPEC_ATTN_BACKEND\",\"draft_sample_method\":\"probabilistic\",$SPEC_REJECTION_CONFIG}"
     )
 fi
 
@@ -694,6 +719,7 @@ VLLM_CMD=(
     "${MAX_MODEL_LEN_ARGS[@]}"
     "${PREFIX_CACHE_ARGS[@]}"
     "${KV_CACHE_DTYPE_ARGS[@]}"
+    "${EAGER_ARGS[@]}"
     "${SPEC_ARGS[@]}"
     "${EVAL_SERVE_ARGS[@]}"
     "${OFFLOAD_ARGS[@]}"

benchmarks/single_node/agentic/kimik3_fp4_mi355x_mtp.sh

@@ -0,0 +1,20 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+# DSpark variant of kimik3_fp4_mi355x.sh. The MI355X launcher routes
+# spec-decoding=mtp rows to this suffix, while the shared base recipe owns the
+# model, KV-offload, AgentX replay, and eval plumbing.
+#
+# These overrides are the DSpark + SimpleCPUOffloadConnector combination
+# validated on gfx950 in PR #2367. Keep them together: widening the batch or
+# re-enabling CUDA graphs reproduced an eight-rank GPU memory access fault.
+export SPEC_DECODE=true
+export KV_CACHE_DTYPE="${KV_CACHE_DTYPE:-auto}"
+export MAX_NUM_SEQS="${MAX_NUM_SEQS:-32}"
+export EVAL_MAX_NUM_SEQS="${EVAL_MAX_NUM_SEQS:-32}"
+export MAX_NUM_BATCHED_TOKENS="${MAX_NUM_BATCHED_TOKENS:-1024}"
+export LANGUAGE_MODEL_ONLY="${LANGUAGE_MODEL_ONLY:-true}"
+export SIMPLE_LAZY_OFFLOAD="${SIMPLE_LAZY_OFFLOAD:-true}"
+export ENFORCE_EAGER="${ENFORCE_EAGER:-true}"
+
+exec "$(dirname "$0")/kimik3_fp4_mi355x.sh" "$@"

configs/amd-master.yaml

@@ -652,6 +652,25 @@ kimik3-fp4-mi355x-vllm-agentic:
       # far the 2234 GiB host tier extends the concurrency ceiling.
       - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, conc-list: [4, 8, 16, 32] }
 
+kimik3-fp4-mi355x-vllm-agentic-mtp:
+  image: vllm/vllm-openai-rocm:kimi-k3
+  model: moonshotai/Kimi-K3
+  model-prefix: kimik3
+  runner: cluster:mi355x-amds
+  precision: fp4
+  framework: vllm
+  multinode: false
+  scenarios:
+    agentic-coding:
+    - dram-utilization: 0.80
+      search-space:
+      # DSpark speculative decoding with Inferact/Kimi-K3-DSpark, seven draft
+      # tokens, probabilistic draft sampling, and the committed thinking-on
+      # golden AL. The _mtp wrapper applies the gfx950 settings validated in
+      # PR #2367 as one unit: eager execution, max-num-batched-tokens 1024,
+      # max-num-seqs 32, bf16 KV, and lazy SimpleCPUOffloadConnector allocation.
+      - { tp: 8, kv-offloading: dram, kv-offload-backend: { name: vllm-simple }, spec-decoding: mtp, conc-list: [1, 2, 4, 8] }
+
 kimik3-fp4-mi355x-vllm-agentic-fp8kvtest:
   image: vllm/vllm-openai-rocm:kimi-k3
   model: moonshotai/Kimi-K3

perf-changelog.yaml

@@ -5273,3 +5273,11 @@
     - "Add 8 checked-in srt-slurm recipes with EAGLE speculative decoding and chat-template inputs"
     - "Image: lmsysorg/sglang:nightly-dev-cu13-20260721-8905cbd4"
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2363
+
+- config-keys:
+    - kimik3-fp4-mi355x-vllm-agentic-mtp
+  description:
+    - "Add the Kimi-K3 MXFP4 MI355X AgentX DSpark variant with Inferact/Kimi-K3-DSpark, seven draft tokens, probabilistic draft sampling, and TRITON_MLA on ROCm."
+    - "Pin throughput to the committed thinking-on golden acceptance length 3.84 with synthetic rejection sampling; eval-only runs retain real block verification."
+    - "Run TP8 with SimpleCPUOffloadConnector at concurrency 1/2/4/8. The _mtp wrapper keeps the gfx950 bring-up settings validated in PR #2367 together: --enforce-eager, --max-num-batched-tokens 1024, --max-num-seqs 32, bf16 KV cache, language-model-only mode, and lazy host allocation."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/2403

utils/changelog_gate_tests/test_run_sweep_gating.py

@@ -409,6 +409,23 @@ def test_e2e_workflow_cannot_dispatch_database_ingest() -> None:
     assert "INFX_FRONTEND_PAT" not in workflow
 
 
+def test_e2e_agentic_eval_preserves_matrix_transport_settings() -> None:
+    workflow = yaml.load(
+        (REPO_ROOT / ".github/workflows/e2e-tests.yml").read_text(),
+        Loader=yaml.BaseLoader,
+    )
+    inputs = workflow["jobs"]["test-sweep-agentic-evals"]["with"]
+
+    assert inputs["kv-offload-backend"] == (
+        "${{ matrix.config['kv-offload-backend'].name }}"
+    )
+    assert inputs["kv-offload-backend-metadata"] == (
+        "${{ matrix.config['kv-offload-backend']"
+        " && toJson(matrix.config['kv-offload-backend']) || '' }}"
+    )
+    assert inputs["spec-decoding"] == "${{ matrix.config.spec-decoding }}"
+
+
 def test_priority_classifier_runs_for_enabled_actions() -> None:
     scenario = {
         **_PR,