[Misc] Reduce supported Punica dtypes (#4304)

This commit is contained in:
Woosuk Kwon
2024-04-23 18:54:33 -07:00
committed by GitHub
parent e4bf860a54
commit 468d761b32
16 changed files with 66 additions and 72 deletions
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_bfloat16, nv_half)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, nv_half, nv_half)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_bfloat16, float, nv_half)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_bfloat16, nv_half)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, nv_half, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, nv_half, float, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_bfloat16, nv_half)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, nv_half, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_bfloat16)
-4
View File
@@ -1,4 +0,0 @@
#include "bgmv_config.h"
#include "bgmv_impl.cuh"
FOR_BGMV_WIDE_NARROW(INST_BGMV_TWOSIDE, float, float, nv_half)
+20
View File
@@ -18,6 +18,26 @@ for input_dtype in DTYPES:
if weight_dtype == "fp32":
# FP32 weights are not supported.
continue
if output_dtype == "fp32":
# LoRA A matrix.
if input_dtype != weight_dtype:
# NOTE(woosuk): While Punica supports the case where the
# input and weight dtypes are different, we only generate
# the kernels the same dtypes to reduce the binary size.
continue
elif input_dtype == "fp32":
# LoRA B matrix.
if output_dtype != weight_dtype:
# NOTE(woosuk): While Punica supports the case where the
# output and weight dtypes are different, we only generate
# the kernels the same dtypes to reduce the binary size.
continue
elif not (input_dtype == output_dtype == weight_dtype):
# NOTE(woosuk): While Punica supports mixed data types for
# input, output, and weight, we only generate the kernels with
# the same data types to reduce the binary size.
continue
kernel_definition = TEMPLATE.format(
input_dtype=DTYPE_MAP[input_dtype],
output_dtype=DTYPE_MAP[output_dtype],
+17
View File
@@ -50,6 +50,23 @@ inline bool launch_bgmv_kernel(out_T *Y, const in_T *X, const W_T *W,
int64_t y_offset, int64_t full_y_size,
int64_t batch_size, int64_t num_layers,
int64_t layer_idx, float scale) {
// NOTE(woosuk): While Punica supports various combinations of input/output
// data types, we limit the supported data types to reduce the binary size.
constexpr bool is_input_float = std::is_same<in_T, float>::value;
constexpr bool is_output_float = std::is_same<out_T, float>::value;
if (is_input_float) {
if (!std::is_same<out_T, W_T>::value) {
return false;
}
} else if (is_output_float) {
if (!std::is_same<in_T, W_T>::value) {
return false;
}
} else if (!(std::is_same<in_T, W_T>::value &&
std::is_same<out_T, W_T>::value)) {
return false;
}
switch (pack_u32(in_features, out_features)) {
#define CASE_ONESIDE(_in_T, _out_T, _W_T, feat_in, feat_out) \
case pack_u32(feat_in, feat_out): \