Gahow Wang
445e491123
third_party/vllm/ now tracked in git for direct patch management.
Based on vLLM v0.18.1 release with one patch applied:
vllm/v1/core/sched/scheduler.py:
Replace fatal assert with graceful skip when KV transfer callback
arrives for an already-aborted request during PD disaggregated serving.
Future vLLM modifications should be made directly in third_party/vllm/
and committed normally. The patches/ directory is kept as documentation
of what changed from upstream.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
32 lines
878 B
Plaintext
32 lines
878 B
Plaintext
#pragma once
|
|
/**
|
|
* __device__ datatypes vectorized by 4
|
|
*/
|
|
|
|
// Include both AMD and NVIDIA fp8 types to avoid circular import
|
|
#include <c10/util/Float8_e4m3fnuz.h>
|
|
#include <c10/util/Float8_e4m3fn.h>
|
|
|
|
namespace vllm {
|
|
|
|
// Vectorization containers
|
|
template <typename scalar_t, size_t vec_size>
|
|
struct __align__(vec_size * sizeof(scalar_t)) vec_n_t {
|
|
scalar_t val[vec_size];
|
|
};
|
|
|
|
template <typename quant_type_t, size_t vec_size>
|
|
struct __align__(vec_size * sizeof(quant_type_t)) q8_n_t {
|
|
static_assert(std::is_same_v<quant_type_t, int8_t> ||
|
|
std::is_same_v<quant_type_t, c10::Float8_e4m3fn> ||
|
|
std::is_same_v<quant_type_t, c10::Float8_e4m3fnuz>);
|
|
quant_type_t val[vec_size];
|
|
};
|
|
|
|
template <typename scalar_t>
|
|
using vec4_t = vec_n_t<scalar_t, 4>;
|
|
template <typename quant_type_t>
|
|
using q8x4_t = q8_n_t<quant_type_t, 4>;
|
|
|
|
} // namespace vllm
|