perf(nvidia): add opt-in Hopper long-context routing

tests/unit_tests/dispatch/test_config_utils.py

@@ -0,0 +1,63 @@
+# Copyright (c) 2026 BAAI. All rights reserved.
+
+from types import SimpleNamespace
+from unittest.mock import patch
+
+from vllm_fl.dispatch.config.utils import get_config_path, load_platform_config
+
+
+def _platform_with_capability(major: int, minor: int = 0):
+    return SimpleNamespace(
+        get_device_capability=lambda: SimpleNamespace(major=major, minor=minor)
+    )
+
+
+def test_hopper_optimization_is_disabled_by_default(monkeypatch):
+    monkeypatch.delenv("VLLM_FL_HOPPER_LONG_CONTEXT_OPT", raising=False)
+    platform = _platform_with_capability(9)
+    with patch("vllm.platforms.current_platform", platform):
+        path = get_config_path("nvidia")
+        config = load_platform_config("nvidia")
+
+    assert path is not None
+    assert path.name == "nvidia.yaml"
+    assert config["op_backends"]["attention_backend"][0] == "flagos"
+    assert "mm" not in config["flagos_blacklist"]
+
+
+def test_hopper_uses_architecture_specific_config_when_enabled(monkeypatch):
+    monkeypatch.setenv("VLLM_FL_HOPPER_LONG_CONTEXT_OPT", "1")
+    platform = _platform_with_capability(9)
+    with patch("vllm.platforms.current_platform", platform):
+        path = get_config_path("nvidia")
+        config = load_platform_config("nvidia")
+
+    assert path is not None
+    assert path.name == "nvidia_hopper.yaml"
+    assert config["op_backends"]["attention_backend"][0] == "vendor"
+    assert {"mm", "mm_out"}.issubset(config["flagos_blacklist"])
+
+
+def test_non_hopper_nvidia_keeps_vendor_wide_config(monkeypatch):
+    monkeypatch.setenv("VLLM_FL_HOPPER_LONG_CONTEXT_OPT", "1")
+    platform = _platform_with_capability(8)
+    with patch("vllm.platforms.current_platform", platform):
+        path = get_config_path("nvidia")
+        config = load_platform_config("nvidia")
+
+    assert path is not None
+    assert path.name == "nvidia.yaml"
+    assert config["op_backends"]["attention_backend"][0] == "flagos"
+    assert "mm" not in config["flagos_blacklist"]
+
+
+def test_capability_probe_failure_falls_back_to_vendor_config(monkeypatch):
+    monkeypatch.setenv("VLLM_FL_HOPPER_LONG_CONTEXT_OPT", "true")
+    platform = SimpleNamespace(
+        get_device_capability=lambda: (_ for _ in ()).throw(RuntimeError("no GPU"))
+    )
+    with patch("vllm.platforms.current_platform", platform):
+        path = get_config_path("nvidia")
+
+    assert path is not None
+    assert path.name == "nvidia.yaml"

vllm_fl/dispatch/README.md

@@ -323,6 +323,7 @@ Environment variables can override specific items from platform config. If not s
 | `VLLM_FL_PREFER` | `flagos` | Preferred backend: `flagos`, `vendor`, `reference` |
 | `VLLM_FL_STRICT` | `0` | Strict mode: `1` = fail on error, `0` = try fallback |
 | `VLLM_FL_PER_OP` | (none) | Per-operator order: `op1=a\|b\|c;op2=x\|y` |
+| `VLLM_FL_HOPPER_LONG_CONTEXT_OPT` | `0` | On NVIDIA Hopper, route attention to FA3 and `mm`/`mm_out` to native CUDA |
 | `VLLM_FL_ALLOW_VENDORS` | (none) | Vendor whitelist, comma-separated |
 | `VLLM_FL_DENY_VENDORS` | (none) | Vendor blacklist, comma-separated |
 
@@ -385,6 +386,9 @@ export VLLM_FL_FLAGOS_WHITELIST="silu_and_mul,rms_norm"
 # Specify per-operator order
 export VLLM_FL_PER_OP="rms_norm=vendor|flagos|reference"
 
+# Opt in to the NVIDIA Hopper long-context routing for an A/B run
+export VLLM_FL_HOPPER_LONG_CONTEXT_OPT=1
+
 # Use completely custom config file
 export VLLM_FL_CONFIG=/path/to/my_config.yaml
 

vllm_fl/dispatch/config/nvidia_hopper.yaml

@@ -0,0 +1,43 @@
+# vLLM-FL Dispatch Configuration for NVIDIA Hopper GPUs
+
+# Keep FlagOS as the default implementation and override only the operators
+# whose Hopper paths have a measured regression against the CUDA vendor stack.
+# This config is opt-in via VLLM_FL_HOPPER_LONG_CONTEXT_OPT=1.
+prefer: flagos
+strict: false
+
+allow_vendors:
+  - cuda
+
+op_backends:
+  # FlashAttention-3 is substantially faster than Triton unified attention
+  # for long-context full-attention prefill on Hopper.
+  attention_backend:
+    - vendor
+    - flagos
+    - reference
+  rms_norm:
+    - flagos
+    - vendor
+    - reference
+  silu_and_mul:
+    - flagos
+    - vendor
+    - reference
+  rotary_embedding:
+    - flagos
+    - vendor
+    - reference
+
+flagos_blacklist:
+  - index_put_
+  - index_put
+  - _index_put_impl_
+  - nonzero
+  - copy_
+  - to_copy
+  - index
+  # FlagGems' Hopper host-TMA mm.out path is slower than the native dense
+  # GEMM for the long-context projection shapes used by vLLM.
+  - mm
+  - mm_out

vllm_fl/dispatch/config/utils.py

@@ -41,6 +41,30 @@
 _CONFIG_DIR = Path(__file__).parent
 
 
+def _get_arch_config_name(platform: str) -> Optional[str]:
+    """Return an architecture-specific config name when one is available."""
+    if platform != "nvidia":
+        return None
+    if os.environ.get("VLLM_FL_HOPPER_LONG_CONTEXT_OPT", "0").lower() not in (
+        "1",
+        "true",
+    ):
+        return None
+
+    try:
+        from vllm.platforms import current_platform
+
+        capability = current_platform.get_device_capability()
+    except Exception:
+        # Device discovery is not guaranteed to be available in API-only or
+        # model-inspection processes. Fall back to the vendor-wide defaults.
+        return None
+
+    if capability is not None and capability.major == 9:
+        return "nvidia_hopper"
+    return None
+
+
 def get_platform_name() -> str:
     """
     Detect the current hardware platform.
@@ -75,6 +99,12 @@ def get_config_path(platform: Optional[str] = None) -> Optional[Path]:
     if platform is None:
         platform = get_platform_name()
 
+    arch_config = _get_arch_config_name(platform)
+    if arch_config is not None:
+        config_file = _CONFIG_DIR / f"{arch_config}.yaml"
+        if config_file.exists():
+            return config_file
+
     # Try platform-specific config
     config_file = _CONFIG_DIR / f"{platform}.yaml"
     if config_file.exists():