review comments

csrc/core/scalar_type.hpp

@@ -14,6 +14,10 @@ namespace vllm {
 //  can be used as a argument for custom operators, helping to simplify these
 //  interfaces.
 //
+//  The type definitions on the Python side can be found in: vllm/_core_ext.pyi
+//  these type definitions should be kept up to date with any Python API changes
+//  here.
+//
 class ScalarType {
  public:
   enum NanRepr : int64_t {
@@ -104,7 +108,7 @@ class ScalarType {
       max_exponent += 1;
     }
 
-    // adjust the exponent to match that off a double
+    // adjust the exponent to match that of a double
     //  for now we assume the exponent bias is the standard 2^(e-1) -1, (where e
     //  is the exponent bits), there is some precedent for non-standard biases,
     //  example `float8_e4m3b11fnuz` here: https://github.com/jax-ml/ml_dtypes
@@ -214,7 +218,9 @@ class ScalarType {
 
   bool operator==(ScalarType const& other) const {
     return mantissa == other.mantissa && exponent == other.exponent &&
-           bias == other.bias && _signed == other._signed;
+           bias == other.bias && _signed == other._signed &&
+           finite_values_only == other.finite_values_only &&
+           nan_repr == other.nan_repr;
   }
 };
 

vllm/_core_ext.py

@@ -10,8 +10,7 @@
 
 if core_C_available:
     try:
-        # ruff: noqa: F401
-        import vllm._core_C
+        import vllm._core_C # ruff: noqa: F401
     except ImportError as e:
         logger.warning("Failed to import from vllm._core_C with %r", e)
 

vllm/_core_ext.pyi

@@ -7,6 +7,9 @@ class ScalarType:
     that torch.dtype currently does not support). It is also cabaable of 
     representing types with a bias, i.e. the stored_value = value + bias, this
     is useful for quantized types (e.g. standard GPTQ 4bit uses a bias of 8).
+    
+    The implementation for this class can be found in csrc/core/scalar_type.hpp,
+    these type definitions should be kept in snyc with that file.
     """
 
     def __init__(self, exponent: int, mantissa: int, bias: int,