minor fixes

csrc/core/scalar_type.hpp

@@ -180,12 +180,12 @@ class ScalarType {
     /*
      * generally follows: https://github.com/jax-ml/ml_dtypes
      * for floating point types (leading f):
-     *  - trailing f: means finite values only (no infinities)
-     *  - trailing n: means nans are supported
-     *  - no-trailing letters: means it follows IEEE 754 conventions
      *  - E_: exponent size
      *  - M_: mantissa size
-     * for integer types:
+     *  - no-trailing letters: means it follows IEEE 754 conventions
+     *  - trailing f: means finite values only (no infinities)
+     *  - trailing n: means nans are supported (non-standard encoding)
+     * for integer types (leading s/u):
      *  - leading s: means signed
      *  - leading u: means unsigned
      *  - number following s/u: number of bits
@@ -337,17 +337,17 @@ using ScalarTypeTorchPtr = c10::intrusive_ptr<ScalarTypeTorch>;
 
 /*
  * generally follows: https://github.com/jax-ml/ml_dtypes
- * for floating point types (leading F):
- *  - trailing f: means finite values only (no infinities)
- *  - trailing n: means nans are supported
- *  - no-trailing letters: means it follows IEEE 754 conventions
+ * for floating point types (leading f):
  *  - E_: exponent size
  *  - M_: mantissa size
- * for integer types:
- *  - leading S: means signed
- *  - leading U: means unsigned
- *  - number following S/U: number of bits
- *  - BX: indicates a non-zero bias of X
+ *  - no-trailing letters: means it follows IEEE 754 conventions
+ *  - trailing f: means finite values only (no infinities)
+ *  - trailing n: means nans are supported (non-standard encoding)
+ * for integer types (leading s/u):
+ *  - leading s: means signed
+ *  - leading u: means unsigned
+ *  - number following s/u: number of bits
+ *  - bX: indicates a non-zero bias of X
  */
 static inline constexpr auto kS4 = ScalarType::s(4);
 static inline constexpr auto kU4 = ScalarType::u(4);

vllm/model_executor/layers/quantization/gptq_marlin_24.py

@@ -285,9 +285,10 @@ def apply(
         size_k = x_2d.shape[1]
         size_n = scales.shape[1]
 
-        output_2d = ops.gptq_marlin_24_gemm(
-            x_2d, qweight, meta, scales, workspace,
-            self.quant_config.quant_type.size_bits, size_m, size_n, size_k)
+        output_2d = ops.gptq_marlin_24_gemm(x_2d, qweight, meta, scales,
+                                            workspace,
+                                            self.quant_config.quant_type,
+                                            size_m, size_n, size_k)
 
         output = output_2d.view(x.shape[:-1] + (output_2d.shape[1], ))
 

vllm/scalar_type.py

@@ -12,12 +12,12 @@ class NanRepr(Enum):
 
 # naming generally follows: https://github.com/jax-ml/ml_dtypes
 # for floating point types (leading f):
-#  - trailing f: means finite values only (no infinities)
-#  - trailing n: means nans are supported
-#  - no-trailing letters: means it follows IEEE 754 conventions
 #  - E_: exponent size
 #  - M_: mantissa size
-# for integer types:
+#  - no-trailing letters: means it follows IEEE 754 conventions
+#  - trailing f: means finite values only (no infinities)
+#  - trailing n: means nans are supported (non-standard encoding)
+# for integer types (leading s/u):
 #  - leading s: means signed
 #  - leading u: means unsigned
 #  - number following s/u: number of bits