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>
65 lines
1.6 KiB
Plaintext
65 lines
1.6 KiB
Plaintext
/*
|
|
Copied from https://github.com/turboderp/exllamav2
|
|
*/
|
|
|
|
#ifndef _compat_cuh
|
|
#define _compat_cuh
|
|
|
|
namespace vllm {
|
|
namespace gptq {
|
|
// atomicAdd for half types, to support CC < 7.x
|
|
|
|
__device__ __forceinline__ void atomicAdd_half(half* address, half val) {
|
|
unsigned int* address_as_ui =
|
|
(unsigned int*)((char*)address - ((size_t)address & 2));
|
|
unsigned int old = *address_as_ui;
|
|
unsigned int assumed;
|
|
|
|
do {
|
|
assumed = old;
|
|
__half_raw hsum;
|
|
hsum.x = (size_t)address & 2 ? (old >> 16) : (old & 0xffff);
|
|
half tmpres = __hadd(hsum, val);
|
|
hsum = __half_raw(tmpres);
|
|
old = (size_t)address & 2 ? (old & 0xffff) | (hsum.x << 16)
|
|
: (old & 0xffff0000) | hsum.x;
|
|
old = atomicCAS(address_as_ui, assumed, old);
|
|
} while (assumed != old);
|
|
}
|
|
|
|
// atomicAdd for half2 types
|
|
|
|
__device__ __forceinline__ void atomicAdd_half2(half2* address, half2 val) {
|
|
unsigned int* address_as_ui = (unsigned int*)address;
|
|
unsigned int old = *address_as_ui;
|
|
unsigned int assumed;
|
|
do {
|
|
assumed = old;
|
|
half2 old_val = *((half2*)&old);
|
|
half2 new_val = __hadd2(old_val, val);
|
|
old = atomicCAS(address_as_ui, assumed, *((unsigned int*)&new_val));
|
|
} while (assumed != old);
|
|
}
|
|
|
|
//
|
|
|
|
#if defined(__CUDA_ARCH__) || defined(USE_ROCM)
|
|
#if __CUDA_ARCH__ < 700 || defined(USE_ROCM)
|
|
|
|
__device__ __forceinline__ void atomicAdd(half* address, half val) {
|
|
atomicAdd_half(address, val);
|
|
}
|
|
|
|
#if __CUDA_ARCH__ < 600 || defined(USE_ROCM)
|
|
__device__ __forceinline__ void atomicAdd(half2* address, half2 val) {
|
|
atomicAdd_half2(address, val);
|
|
}
|
|
#endif
|
|
|
|
#endif
|
|
#endif
|
|
|
|
} // namespace gptq
|
|
} // namespace vllm
|
|
#endif
|