[5/n] Migrate CUTLASS MLA, hadamard, awq, allspark and DSV3 fused a gemm to torch stable ABI (continued) (#42339)

Signed-off-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
Co-authored-by: Mikayla Gawarecki <mikaylagawarecki@gmail.com>
This commit is contained in:
Chris Leonard
2026-05-13 07:24:39 +00:00
committed by GitHub
co-authored by Mikayla Gawarecki
parent 503697c9ce
commit 85b2fecab7
20 changed files with 525 additions and 407 deletions
+26 -18
View File
@@ -1,7 +1,13 @@
#pragma once
// For TORCH_CHECK
#include <torch/library.h>
#include <cstdint>
#include <string>
#include <tuple>
#include <utility>
#include <variant>
// For STD_TORCH_CHECK
#include <torch/headeronly/util/Exception.h>
namespace vllm {
@@ -45,7 +51,7 @@ class ScalarType {
// IEEE 754 compliant floating point type
static constexpr ScalarType float_IEEE754(uint8_t exponent,
uint8_t mantissa) {
TORCH_CHECK(mantissa > 0 && exponent > 0);
STD_TORCH_CHECK(mantissa > 0 && exponent > 0);
return ScalarType(exponent, mantissa, true, 0, false, NAN_IEEE_754);
}
@@ -53,11 +59,12 @@ class ScalarType {
static constexpr ScalarType float_(uint8_t exponent, uint8_t mantissa,
bool finite_values_only,
NanRepr nan_repr) {
TORCH_CHECK(nan_repr < NAN_REPR_ID_MAX, "Invalid NanRepr");
TORCH_CHECK(mantissa > 0 && exponent > 0);
TORCH_CHECK(nan_repr != NAN_IEEE_754,
"use `float_IEEE754` constructor for floating point types that "
"follow IEEE 754 conventions");
STD_TORCH_CHECK(nan_repr < NAN_REPR_ID_MAX, "Invalid NanRepr");
STD_TORCH_CHECK(mantissa > 0 && exponent > 0);
STD_TORCH_CHECK(
nan_repr != NAN_IEEE_754,
"use `float_IEEE754` constructor for floating point types that "
"follow IEEE 754 conventions");
return ScalarType(exponent, mantissa, true, 0, finite_values_only,
nan_repr);
}
@@ -176,8 +183,8 @@ class ScalarType {
private:
double _floating_point_max() const {
TORCH_CHECK(mantissa <= 52 && exponent <= 11,
"Cannot represent max/min as a double for type ", str());
STD_TORCH_CHECK(mantissa <= 52 && exponent <= 11,
"Cannot represent max/min as a double for type ", str());
uint64_t max_mantissa = (uint64_t(1) << mantissa) - 1;
if (nan_repr == NAN_EXTD_RANGE_MAX_MIN) {
@@ -186,8 +193,8 @@ class ScalarType {
uint64_t max_exponent = (uint64_t(1) << exponent) - 2;
if (nan_repr == NAN_EXTD_RANGE_MAX_MIN || nan_repr == NAN_NONE) {
TORCH_CHECK(exponent < 11,
"Cannot represent max/min as a double for type ", str());
STD_TORCH_CHECK(exponent < 11,
"Cannot represent max/min as a double for type ", str());
max_exponent += 1;
}
@@ -216,16 +223,17 @@ class ScalarType {
if (is_floating_point()) {
return {_floating_point_max()};
} else {
TORCH_CHECK(size_bits() < 64 || size_bits() == 64 && is_signed(),
"Cannot represent max as a int64_t");
STD_TORCH_CHECK(size_bits() < 64 || size_bits() == 64 && is_signed(),
"Cannot represent max as a int64_t");
return {(int64_t(1) << mantissa) - 1};
}
}
constexpr std::variant<int64_t, double> _raw_min() const {
if (is_floating_point()) {
TORCH_CHECK(is_signed(),
"We currently assume all floating point types are signed");
STD_TORCH_CHECK(
is_signed(),
"We currently assume all floating point types are signed");
constexpr uint64_t sign_bit_double = (uint64_t(1) << 63);
double max = _floating_point_max();
@@ -233,8 +241,8 @@ class ScalarType {
uint64_t min_raw = max_raw | sign_bit_double;
return {*reinterpret_cast<double*>(&min_raw)};
} else {
TORCH_CHECK(!is_signed() || size_bits() <= 64,
"Cannot represent min as a int64_t");
STD_TORCH_CHECK(!is_signed() || size_bits() <= 64,
"Cannot represent min as a int64_t");
if (is_signed()) {
// set the top bit to 1 (i.e. INT64_MIN) and the rest to 0
// then perform an arithmetic shift right to set all the bits above