From d0251df666ec1651f773badc8109fdc0fe2cbf27 Mon Sep 17 00:00:00 2001
From: Daniel Hanchen <danielhanchen@users.noreply.github.com>
Date: Wed, 11 Feb 2026 15:03:40 +0000
Subject: [PATCH] fix: switch SM100 vLLM workaround to env vars only

---
 unsloth/import_fixes.py | 82 ++++++++---------------------------------
 1 file changed, 15 insertions(+), 67 deletions(-)

diff --git a/unsloth/import_fixes.py b/unsloth/import_fixes.py
index 4a2211f9a0..17a17e7525 100644
--- a/unsloth/import_fixes.py
+++ b/unsloth/import_fixes.py
@@ -999,7 +999,7 @@ def fix_vllm_pdl_blackwell():
         if not torch.cuda.is_available():
             return
 
-        # Scan all GPUs for SM100 - fix applies globally via env var and monkey-patch
+        # Scan all GPUs for SM100 - fix applies globally via env vars
         has_sm100 = False
         sm100_gpu_name = None
         for i in range(torch.cuda.device_count()):
@@ -1014,84 +1014,32 @@ def fix_vllm_pdl_blackwell():
     except Exception:
         return
 
-    # Helper to check if module spec exists
-    def _spec_exists(name):
-        try:
-            return importlib.util.find_spec(name) is not None
-        except (ModuleNotFoundError, ValueError):
-            return False
-
-    # Check if vLLM has the PDL-related modules before doing internet check
-    has_utils = _spec_exists("vllm.lora.ops.triton_ops.utils")
-    has_expand_op = _spec_exists("vllm.lora.ops.triton_ops.lora_expand_op")
-    has_shrink_op = _spec_exists("vllm.lora.ops.triton_ops.lora_shrink_op")
-
-    if not has_utils and not has_expand_op and not has_shrink_op:
-        # Old vLLM version without PDL support - nothing to patch
-        return
-
     # Check if vLLM version includes the fix
     VLLM_PDL_FIX_VERSION = "0.13.2"
     try:
         vllm_version = Version(importlib_version("vllm"))
         if vllm_version > Version(VLLM_PDL_FIX_VERSION):
-            logger.info(
-                f"Unsloth: SM100 ({sm100_gpu_name}) detected but vLLM {vllm_version} "
-                f"should include PDL fix - skipping workaround"
+            logger.debug(
+                f"Unsloth: SM100 ({sm100_gpu_name}) detected with vLLM {vllm_version}. "
+                f"Applying env-only workaround conservatively; user env overrides win."
             )
-            return
     except Exception as e:
         logger.debug(
             f"Unsloth: vLLM version check failed ({e}), applying PDL workaround."
         )
 
-    # Apply the PDL fix
-    os.environ["TRITON_DISABLE_PDL"] = "1"
+    # Keep the workaround minimally intrusive: do not monkey-patch vLLM internals.
+    # Use official env toggles before vLLM import on affected SM100 setups.
+    # "Arch conditional MMA instruction used without targeting appropriate compute capability"
+    # has been observed in this path on Blackwell with older vLLM/torch combinations.
+    os.environ.setdefault("VLLM_LORA_DISABLE_PDL", "1")
+    os.environ.setdefault("TRITON_DISABLE_PDL", "1")
+    os.environ.setdefault("VLLM_USE_FBGEMM", "0")
 
-    def fake_supports_pdl(*args, **kwargs):
-        return False
-
-    patched = []
-
-    # First, patch the source module (utils.py) where supports_pdl is defined.
-    # This is critical because supports_pdl uses @lru_cache - we must clear the
-    # cache to prevent stale cached results from the original function.
-    try:
-        utils_module = importlib.import_module("vllm.lora.ops.triton_ops.utils")
-        if hasattr(utils_module, "supports_pdl"):
-            original_fn = utils_module.supports_pdl
-            if hasattr(original_fn, "cache_clear"):
-                original_fn.cache_clear()
-            utils_module.supports_pdl = fake_supports_pdl
-            patched.append("utils")
-    except (ImportError, ModuleNotFoundError, AttributeError):
-        pass
-
-    # Also patch the consumer modules that import supports_pdl from utils.
-    # This ensures the patched function is used even if the module was already
-    # imported before this fix runs.
-    consumer_modules = {
-        "lora_expand_op": "vllm.lora.ops.triton_ops.lora_expand_op",
-        "lora_shrink_op": "vllm.lora.ops.triton_ops.lora_shrink_op",
-        "fused_moe_lora_op": "vllm.lora.ops.triton_ops.fused_moe_lora_op",
-    }
-    for name, path in consumer_modules.items():
-        try:
-            module = importlib.import_module(path)
-            if hasattr(module, "supports_pdl"):
-                module.supports_pdl = fake_supports_pdl
-                patched.append(name)
-        except (ImportError, ModuleNotFoundError, AttributeError):
-            pass
-
-    if patched:
-        logger.info(
-            f"Unsloth: Applied PDL fix for SM100 ({sm100_gpu_name}) - "
-            f"patched: {', '.join(patched)}"
-        )
-    else:
-        # Just set the env var - vLLM might be an older version without supports_pdl
-        logger.info(f"Unsloth: Set TRITON_DISABLE_PDL=1 for SM100 ({sm100_gpu_name})")
+    logger.info(
+        f"Unsloth: Applied SM100 ({sm100_gpu_name}) vLLM workaround via env vars: "
+        f"VLLM_LORA_DISABLE_PDL=1, TRITON_DISABLE_PDL=1, VLLM_USE_FBGEMM=0"
+    )
 
 
 def patch_openspiel_env_async():