Skip to content
This repository was archived by the owner on Oct 11, 2024. It is now read-only.

Commit 69fa6ed

Browse files
joerunderobertgshaw2-redhat
authored andcommitted
[Kernel] Add punica dimensions for Granite 13b (vllm-project#5559)
Signed-off-by: Joe Runde <Joseph.Runde@ibm.com>
1 parent 18c566f commit 69fa6ed

File tree

2 files changed

+12
-0
lines changed

2 files changed

+12
-0
lines changed

csrc/punica/bgmv/bgmv_config.h

+8
Original file line numberDiff line numberDiff line change
@@ -20,6 +20,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
2020
f(in_T, out_T, W_T, narrow, 1152) \
2121
f(in_T, out_T, W_T, narrow, 1280) \
2222
f(in_T, out_T, W_T, narrow, 1536) \
23+
f(in_T, out_T, W_T, narrow, 1664) \
2324
f(in_T, out_T, W_T, narrow, 1728) \
2425
f(in_T, out_T, W_T, narrow, 1792) \
2526
f(in_T, out_T, W_T, narrow, 2048) \
@@ -36,6 +37,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
3637
f(in_T, out_T, W_T, narrow, 5120) \
3738
f(in_T, out_T, W_T, narrow, 5504) \
3839
f(in_T, out_T, W_T, narrow, 5632) \
40+
f(in_T, out_T, W_T, narrow, 5888) \
3941
f(in_T, out_T, W_T, narrow, 6144) \
4042
f(in_T, out_T, W_T, narrow, 6400) \
4143
f(in_T, out_T, W_T, narrow, 6848) \
@@ -45,6 +47,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
4547
f(in_T, out_T, W_T, narrow, 9216) \
4648
f(in_T, out_T, W_T, narrow, 10240) \
4749
f(in_T, out_T, W_T, narrow, 11008) \
50+
f(in_T, out_T, W_T, narrow, 11264) \
4851
f(in_T, out_T, W_T, narrow, 12288) \
4952
f(in_T, out_T, W_T, narrow, 13696) \
5053
f(in_T, out_T, W_T, narrow, 13824) \
@@ -53,6 +56,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
5356
f(in_T, out_T, W_T, narrow, 16384) \
5457
f(in_T, out_T, W_T, narrow, 20480) \
5558
f(in_T, out_T, W_T, narrow, 22016) \
59+
f(in_T, out_T, W_T, narrow, 22528) \
5660
f(in_T, out_T, W_T, narrow, 24576) \
5761
f(in_T, out_T, W_T, narrow, 27392) \
5862
f(in_T, out_T, W_T, narrow, 27648) \
@@ -91,6 +95,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
9195
f(in_T, out_T, W_T, 1152, narrow) \
9296
f(in_T, out_T, W_T, 1280, narrow) \
9397
f(in_T, out_T, W_T, 1536, narrow) \
98+
f(in_T, out_T, W_T, 1664, narrow) \
9499
f(in_T, out_T, W_T, 1728, narrow) \
95100
f(in_T, out_T, W_T, 1792, narrow) \
96101
f(in_T, out_T, W_T, 2048, narrow) \
@@ -107,6 +112,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
107112
f(in_T, out_T, W_T, 5120, narrow) \
108113
f(in_T, out_T, W_T, 5504, narrow) \
109114
f(in_T, out_T, W_T, 5632, narrow) \
115+
f(in_T, out_T, W_T, 5888, narrow) \
110116
f(in_T, out_T, W_T, 6144, narrow) \
111117
f(in_T, out_T, W_T, 6400, narrow) \
112118
f(in_T, out_T, W_T, 6848, narrow) \
@@ -116,6 +122,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
116122
f(in_T, out_T, W_T, 9216, narrow) \
117123
f(in_T, out_T, W_T, 10240, narrow) \
118124
f(in_T, out_T, W_T, 11008, narrow) \
125+
f(in_T, out_T, W_T, 11264, narrow) \
119126
f(in_T, out_T, W_T, 12288, narrow) \
120127
f(in_T, out_T, W_T, 13696, narrow) \
121128
f(in_T, out_T, W_T, 13824, narrow) \
@@ -124,6 +131,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
124131
f(in_T, out_T, W_T, 16384, narrow) \
125132
f(in_T, out_T, W_T, 20480, narrow) \
126133
f(in_T, out_T, W_T, 22016, narrow) \
134+
f(in_T, out_T, W_T, 22528, narrow) \
127135
f(in_T, out_T, W_T, 24576, narrow) \
128136
f(in_T, out_T, W_T, 27392, narrow) \
129137
f(in_T, out_T, W_T, 27648, narrow) \

tests/lora/test_punica.py

+4
Original file line numberDiff line numberDiff line change
@@ -58,6 +58,7 @@ def _lora_ref_impl(
5858
1152,
5959
1280,
6060
1536,
61+
1664,
6162
2048,
6263
2304,
6364
2560,
@@ -71,6 +72,7 @@ def _lora_ref_impl(
7172
5120,
7273
5504,
7374
5632,
75+
5888,
7476
6144,
7577
6400,
7678
6848,
@@ -80,10 +82,12 @@ def _lora_ref_impl(
8082
9216,
8183
10240,
8284
11008,
85+
11264,
8386
13824,
8487
14336,
8588
15360,
8689
22016,
90+
22528,
8791
24576,
8892
27392,
8993
27648,

0 commit comments

Comments
 (0)