A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 17, 04:15 UTC
Add missing kernel for CodeLlama-34B on A/H100 (no tensor parallelism) when using Multi-LoRA. (#3350)
Or Sharir committed
Mar 13, 2024 at 21:18 UTC
ae0ccb40170d140ded8de99fc905fd8cb0bd409c
2 files changed
+2
-1
csrc/punica/bgmv/bgmv_config.h
+1
@@ -43,6 +43,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
43
f(in_T, out_T, W_T, narrow, 14336) \
44
f(in_T, out_T, W_T, narrow, 16384) \
45
f(in_T, out_T, W_T, narrow, 20480) \
46
+ f(in_T, out_T, W_T, narrow, 22016) \
47
f(in_T, out_T, W_T, narrow, 24576) \
48
f(in_T, out_T, W_T, narrow, 28672) \
49
f(in_T, out_T, W_T, narrow, 32000) \
tests/lora/test_punica.py
+1
-1
@@ -45,7 +45,7 @@ def _lora_ref_impl(
45
H1 = H2 = [
46
128, 256, 512, 1024, 1280, 2048, 2560, 2752, 3072, 3456, 3584, 4096, 5120,
47
5504, 5632, 6144, 6912, 7168, 8192, 9216, 10240, 11008, 13824, 14336,
48
- 24576, 32000, 32256, 32512, 32768, 33024
48
+ 22016, 24576, 32000, 32256, 32512, 32768, 33024
49
]
50
SEED = [0xabcdabcd987]
51