[CI][SM70] Sync #308 to audited latest main

.buildkite/test-amd.yaml

@@ -416,7 +416,7 @@ steps:
   - tests/kernels/helion/
   - vllm/platforms/rocm.py
   commands:
-  - pip install helion==1.0.0
+  - pip install helion==1.4.0
   - pytest -v -s kernels/helion/
 
 - label: Kernels Mamba Test # TBD

.buildkite/test_areas/kernels.yaml

@@ -223,7 +223,7 @@ steps:
   - vllm/utils/import_utils.py
   - tests/kernels/helion/
   commands:
-    - pip install helion==1.0.0
+    - pip install helion==1.4.0
     - pytest -v -s kernels/helion/
 
 

.pre-commit-config.yaml

@@ -4,7 +4,7 @@ default_install_hook_types:
 default_stages:
   - pre-commit # Run locally
   - manual # Run in CI
-exclude: '(vllm/third_party/.*|flash_qla/.*)'
+exclude: '(vllm/third_party/.*|flash_qla/.*|csrc/sm70_turbomind/lmdeploy/.*)'
 repos:
 - repo: https://github.com/astral-sh/ruff-pre-commit
   rev: v0.14.0
@@ -35,6 +35,11 @@ repos:
   rev: v1.7.7
   hooks:
   - id: actionlint
+- repo: https://github.com/shellcheck-py/shellcheck-py
+  rev: 745eface02aef23e168a8afb6b5737818efbea95 # v0.11.0.1
+  hooks:
+  - id: shellcheck
+    args: ["-s", "bash", "--format=gcc", "--severity=error"]
 - repo: https://github.com/astral-sh/uv-pre-commit
   rev: 0.11.1
   hooks:
@@ -177,11 +182,6 @@ repos:
     entry: python tools/pre_commit/mypy.py 1 "3.13"
     <<: *mypy_common
     stages: [manual] # Only run in CI
-  - id: shellcheck
-    name: Lint shell scripts
-    entry: tools/pre_commit/shellcheck.sh
-    language: script
-    types: [shell]
   - id: png-lint
     name: Lint PNG exports from excalidraw
     entry: tools/pre_commit/png-lint.sh

CMakeLists.txt

@@ -397,6 +397,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA")
       "${VLLM_SM70_TURBOMIND_LMDEPLOY_ROOT}/src/turbomind/kernels/gemm/kernel/sm70_884_16.cu"
       "${VLLM_SM70_TURBOMIND_ROOT}/ops/tm_registry_sm70.cu"
       "${VLLM_SM70_TURBOMIND_ROOT}/ops/fp8_qpn8_sm70.cu"
+      "${VLLM_SM70_TURBOMIND_ROOT}/ops/mxfp4_qpn_m1_sm70.cu"
       "${VLLM_SM70_TURBOMIND_ROOT}/ops/nvfp4_qpn4_sm70.cu"
       "${VLLM_SM70_TURBOMIND_ROOT}/ops/qwen38_prefill_cutlass.cu"
       "${VLLM_SM70_TURBOMIND_ROOT}/ops/nvfp4_qpn2_sm70.cu"
@@ -729,6 +730,7 @@ if(VLLM_GPU_LANG STREQUAL "CUDA" AND SM70_TURBOMIND_ARCHS)
     SOURCES ${VLLM_SM70_SAMPLER_FRAGMENT_SRCS}
     COMPILE_FLAGS ${VLLM_GPU_FLAGS}
     ARCHITECTURES ${VLLM_GPU_ARCHES}
+    USE_SABI 3
     WITH_SOABI)
   set_target_properties(_sm70_sampler_C PROPERTIES
     CXX_STANDARD 17

README.md

@@ -349,9 +349,9 @@ python -m pip install -e . --no-build-isolation
 
 请使用微信扫描下方二维码加入群组：
 
-![1Cat-vLLM 微信交流群二维码](docs/assets/wechat-group-qr-7.png)
+<img width="966" height="1518" alt="e52eca541f6c191928963c97ac1b38e5" src="https://github.com/user-attachments/assets/73be9e1e-41c3-49d7-9d4e-db4ce97cce9b" />
 
-> 提示：微信群二维码通常 7 天内有效。若扫描失败或提示过期，请重新打开本页查看最新图片，或关注仓库更新。
+> 提示：微信群二维码通常 7 天内有效。若扫描失败或提示过期，请微信搜索微信号添加：YM_isi。
 
 ## Repository Notes
 

benchmarks/benchmark_dspark_api.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 """Issue one reproducible OpenAI chat request for DSpark validation."""
 

benchmarks/benchmark_dspark_stream.py

@@ -1,4 +1,5 @@
 # SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
 
 """Measure one streamed completions request with separate TTFT and TPOT."""
 

benchmarks/benchmark_sm70_decode.py

@@ -52,20 +52,42 @@ def _sm70_fa2_d256_prefill_status(torch: Any) -> dict[str, Any]:
         "sm70_d256_splitd_n32_paged_fwd",
     )
     optional_ops = ("sm70_d256_splitd_n32_dense_splitkv3_fwd",)
-    try:
-        importlib.import_module("vllm.vllm_flash_attn.flash_attn_interface")
-    except (AttributeError, ImportError, RuntimeError) as exc:
-        return {
-            "available": False,
-            "error": f"{type(exc).__name__}: {exc}",
-            "extension_file": None,
-            "extension_realpath": None,
-            "required_ops": {name: False for name in required_ops},
-            "optional_ops": {name: False for name in optional_ops},
-        }
-
     extension_file = _module_file("vllm.vllm_flash_attn._vllm_fa2_C")
     namespace = getattr(torch.ops, "_vllm_fa2_C", None)
+    import_error: Exception | None = None
+    if namespace is None or not all(hasattr(namespace, name) for name in required_ops):
+        try:
+            importlib.import_module("vllm.vllm_flash_attn.flash_attn_interface")
+        except (AttributeError, ImportError, RuntimeError) as exc:
+            import_error = exc
+        namespace = getattr(torch.ops, "_vllm_fa2_C", None)
+    if namespace is None or not all(hasattr(namespace, name) for name in required_ops):
+        library_path = os.getenv("VLLM_SM70_FA2_D256_LIBRARY")
+        if library_path is not None:
+            try:
+                torch.ops.load_library(library_path)
+            except (OSError, RuntimeError) as load_exc:
+                return {
+                    "available": False,
+                    "error": f"{type(load_exc).__name__}: {load_exc}",
+                    "extension_file": library_path,
+                    "extension_realpath": str(Path(library_path).resolve()),
+                    "required_ops": {name: False for name in required_ops},
+                    "optional_ops": {name: False for name in optional_ops},
+                }
+            extension_file = library_path
+        elif import_error is not None:
+            return {
+                "available": False,
+                "error": f"{type(import_error).__name__}: {import_error}",
+                "extension_file": None,
+                "extension_realpath": None,
+                "required_ops": {name: False for name in required_ops},
+                "optional_ops": {name: False for name in optional_ops},
+            }
+        namespace = getattr(torch.ops, "_vllm_fa2_C", None)
+    if extension_file is None:
+        extension_file = os.getenv("VLLM_SM70_FA2_D256_LIBRARY")
     required_status = {
         name: namespace is not None and hasattr(namespace, name)
         for name in required_ops
@@ -169,6 +191,7 @@ def _tracked_env() -> dict[str, str]:
         "VLLM_FLASH_",
         "FLASH_ATTN",
         "FLA_",
+        "PREFIX_",
         "CUDA_MODULE_LOADING",
         "TORCH_CUDA_ARCH_LIST",
         "TORCHINDUCTOR_",
@@ -1172,6 +1195,8 @@ def _sm70_gdn_fla_policy() -> dict[str, Any]:
 
 def _sm70_moe_policy() -> dict[str, Any]:
     add_allreduce = _env_bool("VLLM_SM70_MOE_ADD_ALLREDUCE", False)
+    qwen38_router_topk = _env_bool("VLLM_SM70_QWEN38_ROUTER_TOPK", True)
+    qwen38_nvfp4_qpn_m1 = _env_bool("VLLM_SM70_NVFP4_QWEN38_MOE_QPN_M1_DECODE", True)
     single_token_fastpath = _env_bool("VLLM_SM70_MOE_SINGLE_TOKEN_FASTPATH", False)
     single_token_permute = _env_bool(
         "VLLM_SM70_MOE_SINGLE_TOKEN_PERMUTE_FASTPATH", False
@@ -1203,6 +1228,12 @@ def _sm70_moe_policy() -> dict[str, Any]:
     return {
         "VLLM_SM70_MOE_ADD_ALLREDUCE": os.environ.get("VLLM_SM70_MOE_ADD_ALLREDUCE"),
         "moe_add_allreduce_effective": add_allreduce,
+        "VLLM_SM70_QWEN38_ROUTER_TOPK": os.environ.get("VLLM_SM70_QWEN38_ROUTER_TOPK"),
+        "qwen38_router_topk_effective": qwen38_router_topk,
+        "VLLM_SM70_NVFP4_QWEN38_MOE_QPN_M1_DECODE": os.environ.get(
+            "VLLM_SM70_NVFP4_QWEN38_MOE_QPN_M1_DECODE"
+        ),
+        "qwen38_nvfp4_qpn_m1_effective": qwen38_nvfp4_qpn_m1,
         "VLLM_SM70_MOE_SINGLE_TOKEN_FASTPATH": os.environ.get(
             "VLLM_SM70_MOE_SINGLE_TOKEN_FASTPATH"
         ),
@@ -1249,8 +1280,16 @@ def _sm70_moe_policy() -> dict[str, Any]:
         ),
         "profile_trace_effective": profile_trace,
         "expected_default": False,
+        "qwen38_router_topk_expected_default": True,
+        "qwen38_nvfp4_qpn_m1_expected_default": True,
         "single_token_unpermute_expected_default": True,
         "route_hit_oracle": (
+            "The exact Qwen3.8 E512/K10 lane requires "
+            "qwen38_router_topk_effective=true and the "
+            "`SM70 Qwen3.8 E512/K10 router top-k path enabled` marker. "
+            "The accepted routed-expert lane additionally requires "
+            "qwen38_nvfp4_qpn_m1_effective=true and the `SM70 Qwen3.8 "
+            "NVFP4 direct QPN-M1 expert path enabled` marker. "
             "For CUDA-graph fast-path acceptance, require the C++ trace "
             "`SM70 custom all_reduce_sum2 op reached ... capture=active`; "
             "the Python MoERunner candidate log alone is not a custom-op hit. "
@@ -1267,6 +1306,7 @@ def _sm70_moe_policy() -> dict[str, Any]:
 
 
 def _sm70_sampling_policy() -> dict[str, Any]:
+    use_v2_model_runner = _env_bool("VLLM_USE_V2_MODEL_RUNNER", False)
     flash_0dot3_compile = _env_bool(
         "VLLM_SM70_FLASH_V100_0DOT3_COMPILE_GRAPH",
         False,
@@ -1279,6 +1319,8 @@ def _sm70_sampling_policy() -> dict[str, Any]:
     full_lm_head = _env_bool("VLLM_SM70_ENABLE_LM_HEAD_FASTPATH", False)
     dense_f16 = _env_bool("VLLM_SM70_ENABLE_DENSE_F16_FASTPATH", False)
     return {
+        "VLLM_USE_V2_MODEL_RUNNER": os.environ.get("VLLM_USE_V2_MODEL_RUNNER"),
+        "use_v2_model_runner_effective": use_v2_model_runner,
         "VLLM_SM70_GREEDY_TOKEN_FASTPATH": os.environ.get(
             "VLLM_SM70_GREEDY_TOKEN_FASTPATH"
         ),
@@ -1319,14 +1361,25 @@ def _sm70_sampling_policy() -> dict[str, Any]:
             and not full_lm_head
             and not custom_top1_ar
         ),
+        "v2_greedy_tp_local_pair_policy": (
+            use_v2_model_runner
+            and greedy_fastpath
+            and not lm_head_top1
+            and not lm_head_top1_tc
+            and not full_lm_head
+            and not custom_top1_ar
+        ),
         "route_hit_oracle": (
             "Accepted default greedy decode requires "
             "`SM70 LM head top1 layout prepared` plus "
             "`SM70 fused LM head top1 path enabled`; "
             "The SM70 Flash-V100 compile-graph quality lane instead requires "
             "VLLM_SM70_LM_HEAD_TOP1=0 and "
             "compile_graph_local_logits_top1_policy=true, preserving greedy "
-            "pair-gather while avoiding the fused top1 epilogue. "
+            "pair-gather while avoiding the fused top1 epilogue. The MRv2 "
+            "single-token lane additionally requires "
+            "v2_greedy_tp_local_pair_policy=true and the "
+            "`SM70 MRv2 greedy TP-local pair path enabled` log marker. "
             "`SM70 dense fp16 fast path enabled for LM head` is expected only "
             "when the explicit full LM-head fast path gate is enabled. "
             "The 0.0.3 LM-head baseline lane requires full LM-head fast path "