Update B200 MiniMax-M3 vLLM and remove the KV offload patch

benchmarks/single_node/agentic/minimaxm3_fp4_b200_mtp.sh

@@ -126,7 +126,6 @@ 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))
     export VLLM_USE_SIMPLE_KV_OFFLOAD=1
     OFFLOAD_CONFIG=$(printf \

configs/nvidia-master.yaml

@@ -7983,7 +7983,7 @@ minimaxm3-fp4-b300-trtllm-agentic-mtp:
 # the same 3 TB AgentX ceiling before the proportional-GPU rule is applied.
 # GPU-resident points receive a zero budget.
 minimaxm3-fp4-b200-vllm-agentic-mtp:
-  image: vllm/vllm-openai:nightly-1dc464d42681d22f38caf1fdc1eb632dc4421c45
+  image: vllm/vllm-openai:nightly-cd10ed6f9f6b37a8ace9cf380007e66fe12ec0c3
   model: nvidia/MiniMax-M3-NVFP4
   model-prefix: minimaxm3
   runner: cluster:b200-nscale

perf-changelog.yaml

@@ -7807,3 +7807,11 @@
     - "Use FLASH_ATTN for the target model, select the Humming MXFP8 dense-linear backend, and enable Marlin atomic reduction for the MXFP8 MoE path."
     - "Use 0.95 GPU memory utilization for resident serving while retaining 0.90 for Mooncake DRAM offload, and remove the Mooncake concurrency-14 point."
   pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3108
+
+- config-keys:
+    - minimaxm3-fp4-b200-vllm-agentic-mtp
+  scenario-type:
+    - agentic-coding
+  description:
+    - "Update B200 MiniMax-M3 AgentX to the latest vLLM image with upstream mixed-page KV offload support and remove the runtime offload patch."
+  pr-link: https://github.com/SemiAnalysisAI/InferenceX/pull/3131

runners/test_slurm_utils.py

@@ -13,7 +13,6 @@
 PATCH_SRT_EVAL = REPO_ROOT / "runners" / "patch_srt_eval_dispatch.py"
 PATCH_SRT_DP_RANKS = REPO_ROOT / "runners" / "patch_srt_vllm_dp_ranks.py"
 PATCH_TRTLLM_CHAT_STORE = REPO_ROOT / "runners" / "patch_trtllm_chat_store.py"
-PATCH_VLLM_SIMPLE_KV = REPO_ROOT / "runners" / "patch_vllm_simple_kv_offload.py"
 INJECT_ACCEPTANCE = REPO_ROOT / "runners" / "inject_synthetic_acceptance.py"
 
 
@@ -393,52 +392,6 @@ def test_patch_trtllm_chat_store_rejects_unknown_source(tmp_path: Path) -> None:
     assert result.returncode == 1
     assert protocol.read_text() == "unsupported protocol\n"
 
-def test_patch_vllm_simple_kv_offload_is_idempotent_and_preserves_surrounding_code(
-    tmp_path: Path,
-) -> None:
-    symbols = runpy.run_path(str(PATCH_VLLM_SIMPLE_KV))
-    worker = tmp_path / "worker.py"
-    original = f"prefix\n{symbols['OLD_SETUP']}{symbols['OLD_LOOP']}suffix\n"
-    worker.write_text(original)
-
-    first = subprocess.run(
-        ["python3", str(PATCH_VLLM_SIMPLE_KV), str(worker)],
-        check=False,
-        capture_output=True,
-        text=True,
-    )
-    patched = worker.read_text()
-    second = subprocess.run(
-        ["python3", str(PATCH_VLLM_SIMPLE_KV), str(worker)],
-        check=False,
-        capture_output=True,
-        text=True,
-    )
-
-    assert first.returncode == 0, first.stderr
-    assert second.returncode == 0, second.stderr
-    assert patched != original
-    assert patched.startswith("prefix\n") and patched.endswith("suffix\n")
-    assert worker.read_text() == patched
-
-
-def test_patch_vllm_simple_kv_offload_rejects_unknown_source(
-    tmp_path: Path,
-) -> None:
-    worker = tmp_path / "worker.py"
-    worker.write_text("unsupported worker\n")
-
-    result = subprocess.run(
-        ["python3", str(PATCH_VLLM_SIMPLE_KV), str(worker)],
-        check=False,
-        capture_output=True,
-        text=True,
-    )
-
-    assert result.returncode == 1
-    assert worker.read_text() == "unsupported worker\n"
-
-
 def test_patch_srt_eval_dispatch_preflights_before_writing(tmp_path: Path) -> None:
     do_sweep = tmp_path / "src/srtctl/cli/do_sweep.py"
     eval_script = tmp_path / "src/srtctl/benchmarks/scripts/lm-eval/bench.sh"