fix: handle vendor-prefixed GGUF quant types (e.g., UD-Q4_K_XL)

tests/transformers_utils/test_utils.py

@@ -8,6 +8,7 @@
 from vllm.transformers_utils.gguf_utils import (
     is_gguf,
     is_remote_gguf,
+    is_valid_gguf_quant_type,
     split_remote_gguf,
 )
 from vllm.transformers_utils.utils import (
@@ -119,6 +120,16 @@ def test_is_remote_gguf_with_cloud_storage(self):
         assert not is_remote_gguf("s3://repo/model:Q4_K")
         assert not is_remote_gguf("gs://repo/model:Q8_0")
 
+    def test_vendor_prefixed_remote_gguf(self):
+        """Remote GGUF with vendor-prefixed quant types should be detected."""
+        assert is_remote_gguf("unsloth/Qwen3.5-35B-A3B-GGUF:UD-Q4_K_XL")
+        assert is_remote_gguf("repo/model:UD-Q4_K_M")
+        assert is_remote_gguf("repo/model:UD-Q3_K_S")
+        assert is_remote_gguf("repo/model:UD-F16")
+
+        # Invalid vendor-prefixed should still fail
+        assert not is_remote_gguf("repo/model:UD-INVALID")
+
 
 class TestSplitRemoteGGUF:
     """Test split_remote_gguf utility function."""
@@ -167,6 +178,14 @@ def test_split_remote_gguf_invalid(self):
         with pytest.raises(ValueError, match="Wrong GGUF model"):
             split_remote_gguf("s3://bucket/repo/model:Q2_K")
 
+    def test_split_vendor_prefixed_remote_gguf(self):
+        """split_remote_gguf should preserve full quant string including prefix."""
+        repo_id, quant_type = split_remote_gguf(
+            "unsloth/Qwen3.5-35B-A3B-GGUF:UD-Q4_K_XL"
+        )
+        assert repo_id == "unsloth/Qwen3.5-35B-A3B-GGUF"
+        assert quant_type == "UD-Q4_K_XL"
+
 
 class TestIsGGUF:
     """Test is_gguf utility function."""
@@ -218,3 +237,41 @@ def test_is_gguf_edge_cases(self):
         # Cloud storage
         assert not is_gguf("s3://bucket/repo/model:IQ1_S")
         assert not is_gguf("gs://bucket/repo/model:Q2_K")
+
+    def test_vendor_prefixed_is_gguf(self):
+        """is_gguf should recognize vendor-prefixed remote GGUF."""
+        assert is_gguf("unsloth/Qwen3.5-35B-A3B-GGUF:UD-Q4_K_XL")
+        assert is_gguf("repo/model:UD-Q4_K_M")
+        assert not is_gguf("repo/model:UD-INVALID")
+
+
+class TestIsValidGGUFQuantType:
+    """Test is_valid_gguf_quant_type utility function."""
+
+    def test_vendor_prefixed_quant_types(self):
+        """Vendor-prefixed quant types like UD-Q4_K_XL should be valid."""
+        # Unsloth Dynamic (UD-) prefix with size suffix
+        assert is_valid_gguf_quant_type("UD-Q4_K_XL")
+        assert is_valid_gguf_quant_type("UD-Q4_K_M")
+        assert is_valid_gguf_quant_type("UD-Q3_K_S")
+        assert is_valid_gguf_quant_type("UD-Q5_K_L")
+        assert is_valid_gguf_quant_type("UD-Q6_K_XXS")
+
+        # UD- prefix with exact GGML types (no size suffix)
+        assert is_valid_gguf_quant_type("UD-Q4_K")
+        assert is_valid_gguf_quant_type("UD-Q4_0")
+        assert is_valid_gguf_quant_type("UD-F16")
+        assert is_valid_gguf_quant_type("UD-BF16")
+        assert is_valid_gguf_quant_type("UD-IQ1_S")
+
+        # Other hypothetical vendor prefixes should also work
+        assert is_valid_gguf_quant_type("XX-Q4_K_M")
+
+        # Invalid base type after valid prefix should fail
+        assert not is_valid_gguf_quant_type("UD-INVALID")
+        assert not is_valid_gguf_quant_type("UD-Q9_K_M")
+        assert not is_valid_gguf_quant_type("UD-")
+
+        # Empty prefix should fail (e.g., "-Q4_K" is not a valid quant type)
+        assert not is_valid_gguf_quant_type("-Q4_K")
+        assert not is_valid_gguf_quant_type("-")

vllm/transformers_utils/gguf_utils.py

@@ -54,26 +54,51 @@ def is_remote_gguf(model: str | Path) -> bool:
 _GGUF_QUANT_SUFFIXES = ("_M", "_S", "_L", "_XL", "_XS", "_XXS")
 
 
-def is_valid_gguf_quant_type(gguf_quant_type: str) -> bool:
-    """Check if the quant type is a valid GGUF quant type.
+def _is_base_gguf_quant_type(quant_type: str) -> bool:
+    """Check if quant_type matches a GGMLQuantizationType, with optional
+    size suffix.
 
-    Supports both exact GGML quant types (e.g., Q4_K, IQ1_S) and
-    extended naming conventions (e.g., Q4_K_M, Q3_K_S, Q5_K_L).
+    Checks for exact enum match first, then tries stripping known suffixes
+    (e.g., Q4_K_M -> Q4_K).
     """
-    # Check for exact match first
-    if getattr(GGMLQuantizationType, gguf_quant_type, None) is not None:
+    if getattr(GGMLQuantizationType, quant_type, None) is not None:
         return True
 
-    # Check for extended naming conventions (e.g., Q4_K_M -> Q4_K)
     for suffix in _GGUF_QUANT_SUFFIXES:
-        if gguf_quant_type.endswith(suffix):
-            base_type = gguf_quant_type[: -len(suffix)]
+        if quant_type.endswith(suffix):
+            base_type = quant_type[: -len(suffix)]
             if getattr(GGMLQuantizationType, base_type, None) is not None:
                 return True
 
     return False
 
 
+def is_valid_gguf_quant_type(gguf_quant_type: str) -> bool:
+    """Check if the quant type is a valid GGUF quant type.
+
+    Supports exact GGML quant types (e.g., Q4_K, IQ1_S), extended naming
+    conventions with size suffixes (e.g., Q4_K_M, Q3_K_S, Q5_K_L), and
+    vendor-prefixed quant types (e.g., UD-Q4_K_XL) where the prefix is
+    separated by a hyphen.
+
+    Vendor prefixes are stripped because GGML quant type names never
+    contain hyphens — any hyphen indicates a vendor prefix (e.g., "UD-"
+    for Unsloth Dynamic quantization).
+    """
+    if _is_base_gguf_quant_type(gguf_quant_type):
+        return True
+
+    # Try stripping vendor prefix (e.g., "UD-Q4_K_XL" -> "Q4_K_XL").
+    # GGML quant type names never contain hyphens, so a hyphen indicates
+    # a vendor prefix.
+    if "-" in gguf_quant_type:
+        prefix, remainder = gguf_quant_type.rsplit("-", 1)
+        if prefix and remainder:
+            return _is_base_gguf_quant_type(remainder)
+
+    return False
+
+
 def split_remote_gguf(model: str | Path) -> tuple[str, str]:
     """Split the model into repo_id and quant type."""
     model = str(model)
@@ -84,7 +109,8 @@ def split_remote_gguf(model: str | Path) -> tuple[str, str]:
         f"Wrong GGUF model or invalid GGUF quant type: {model}.\n"
         "- It should be in repo_id:quant_type format.\n"
         f"- Valid base quant types: {GGMLQuantizationType._member_names_}\n"
-        f"- Extended suffixes also supported: {_GGUF_QUANT_SUFFIXES}",
+        f"- Extended suffixes also supported: {_GGUF_QUANT_SUFFIXES}\n"
+        "- Vendor-prefixed types also supported (e.g., UD-Q4_K_XL)",
     )