Fix attention backend auto-detection for consumer Blackwell (sm120)

python/minisgl/engine/engine.py

@@ -11,7 +11,14 @@
 from minisgl.layers import set_rope_device
 from minisgl.models import create_model, load_weight
 from minisgl.moe import create_moe_backend
-from minisgl.utils import div_even, init_logger, is_sm90_supported, is_sm100_supported, torch_dtype
+from minisgl.utils import (
+    div_even,
+    init_logger,
+    is_sm90_supported,
+    is_sm100_supported,
+    is_sm120_supported,
+    torch_dtype,
+)
 
 from .config import EngineConfig
 from .graph import GraphRunner, get_free_memory, mem_GB
@@ -220,7 +227,14 @@ def override(attr: str, value: Any):  # this is dangerous, use with caution
         object.__setattr__(config, attr, value)
 
     if config.attention_backend == "auto":
-        backend = "trtllm" if is_sm100_supported() else ("fa,fi" if is_sm90_supported() else "fi")
+        if is_sm120_supported():
+            backend = "fi"
+        elif is_sm100_supported():
+            backend = "trtllm"
+        elif is_sm90_supported():
+            backend = "fa,fi"
+        else:
+            backend = "fi"
         override("attention_backend", backend)
         logger.info_rank0(f"Auto-selected attention backend: {config.attention_backend}")
 

python/minisgl/utils/__init__.py

@@ -1,4 +1,4 @@
-from .arch import is_arch_supported, is_sm90_supported, is_sm100_supported
+from .arch import is_arch_supported, is_sm90_supported, is_sm100_supported, is_sm120_supported
 from .hf import cached_load_hf_config, download_hf_weight, load_tokenizer
 from .logger import init_logger
 from .misc import UNSET, Unset, align_ceil, align_down, call_if_main, div_ceil, div_even
@@ -21,6 +21,7 @@
     "is_arch_supported",
     "is_sm90_supported",
     "is_sm100_supported",
+    "is_sm120_supported",
     "call_if_main",
     "div_even",
     "div_ceil",

python/minisgl/utils/arch.py

@@ -27,3 +27,9 @@ def is_sm90_supported() -> bool:
 
 def is_sm100_supported() -> bool:
     return is_arch_supported(10, 0)
+
+
+def is_sm120_supported() -> bool:
+    # Consumer Blackwell (sm120) numerically passes the sm100 check but doesn't support
+    # trtllm-gen/FA4 kernels, which target datacenter Blackwell (sm100/B200) only.
+    return is_arch_supported(12, 0)