mirror of
https://github.com/zenlm/engine.git
synced 2026-10-30 15:00:05 +00:00
Eric BuehlerandGitHub
863f33de71
* Add chunking to naive attn * Add it to cublaslt as well * Even better! * Clippy * Remove prompt chunksize apis * Prompt chunksize logging * Format * Determine based on mask type * Max seq len in device map * Max seq len in device map * Format
50 lines
1.8 KiB
Plaintext
50 lines
1.8 KiB
Plaintext
#include <cassert>
|
|
#include <cstdint>
|
|
#include <cuda.h>
|
|
|
|
extern "C" void launch_dequant_fp8_blockwise_kernel_f32(
|
|
const uint8_t *d_weight, const float *d_scale, float *d_output,
|
|
int weight_height, int weight_width, int weight_row_stride,
|
|
int scale_stride, int weight_block_size_y, int weight_block_size_x,
|
|
cudaStream_t stream) {
|
|
assert(false);
|
|
}
|
|
|
|
extern "C" void launch_dequant_fp8_blockwise_kernel_f16(
|
|
const uint8_t *d_weight, const float *d_scale, uint16_t *d_output,
|
|
int weight_height, int weight_width, int weight_row_stride,
|
|
int scale_stride, int weight_block_size_y, int weight_block_size_x,
|
|
cudaStream_t stream) {
|
|
assert(false);
|
|
}
|
|
|
|
extern "C" void launch_dequant_fp8_blockwise_kernel_bf16(
|
|
const uint8_t *d_weight, const float *d_scale, uint16_t *d_output,
|
|
int weight_height, int weight_width, int weight_row_stride,
|
|
int scale_stride, int weight_block_size_y, int weight_block_size_x,
|
|
cudaStream_t stream) {
|
|
assert(false);
|
|
}
|
|
|
|
extern "C" void launch_quant_fp8_blockwise_kernel_f32(
|
|
const float *d_input, uint8_t *d_weight, float *d_scale, int weight_height,
|
|
int weight_width, int weight_row_stride, int scale_stride,
|
|
int weight_block_size_y, int weight_block_size_x, cudaStream_t stream) {
|
|
assert(false);
|
|
}
|
|
|
|
extern "C" void launch_quant_fp8_blockwise_kernel_f16(
|
|
const uint16_t *d_input, uint8_t *d_weight, float *d_scale,
|
|
int weight_height, int weight_width, int weight_row_stride,
|
|
int scale_stride, int weight_block_size_y, int weight_block_size_x,
|
|
cudaStream_t stream) {
|
|
assert(false);
|
|
}
|
|
|
|
extern "C" void launch_quant_fp8_blockwise_kernel_bf16(
|
|
const uint16_t *d_input, uint8_t *d_weight, float *d_scale,
|
|
int weight_height, int weight_width, int weight_row_stride,
|
|
int scale_stride, int weight_block_size_y, int weight_block_size_x,
|
|
cudaStream_t stream) {
|
|
assert(false);
|
|
} |