Add missing kernel for CodeLlama-34B on A/H100 (no tensor parallelism) when...

Add missing kernel for CodeLlama-34B on A/H100 (no tensor parallelism) when using Multi-LoRA. (#3350)

csrc/punica/bgmv/bgmv_config.h

@@ -43,6 +43,7 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 14336) \
     f(in_T, out_T, W_T, narrow, 16384) \
     f(in_T, out_T, W_T, narrow, 20480) \
+    f(in_T, out_T, W_T, narrow, 22016) \
     f(in_T, out_T, W_T, narrow, 24576) \
     f(in_T, out_T, W_T, narrow, 28672) \
     f(in_T, out_T, W_T, narrow, 32000) \

tests/lora/test_punica.py

@@ -45,7 +45,7 @@ def _lora_ref_impl(
 H1 = H2 = [
     128, 256, 512, 1024, 1280, 2048, 2560, 2752, 3072, 3456, 3584, 4096, 5120,
     5504, 5632, 6144, 6912, 7168, 8192, 9216, 10240, 11008, 13824, 14336,
-    24576, 32000, 32256, 32512, 32768, 33024
+    22016, 24576, 32000, 32256, 32512, 32768, 33024
 ]
 SEED = [0xabcdabcd987]