diff --git a/csrc/punica/bgmv/bgmv_config.h b/csrc/punica/bgmv/bgmv_config.h
index c38db2dcd2c4d..cb6694b3036e9 100755
--- a/csrc/punica/bgmv/bgmv_config.h
+++ b/csrc/punica/bgmv/bgmv_config.h
@@ -16,15 +16,20 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 512) \
     f(in_T, out_T, W_T, narrow, 640) \
     f(in_T, out_T, W_T, narrow, 768) \
+    f(in_T, out_T, W_T, narrow, 896) \
     f(in_T, out_T, W_T, narrow, 1024) \
     f(in_T, out_T, W_T, narrow, 1152) \
+    f(in_T, out_T, W_T, narrow, 1216) \
     f(in_T, out_T, W_T, narrow, 1280) \
     f(in_T, out_T, W_T, narrow, 1536) \
     f(in_T, out_T, W_T, narrow, 1664) \
     f(in_T, out_T, W_T, narrow, 1728) \
     f(in_T, out_T, W_T, narrow, 1792) \
     f(in_T, out_T, W_T, narrow, 2048) \
+    f(in_T, out_T, W_T, narrow, 2240) \
     f(in_T, out_T, W_T, narrow, 2304) \
+    f(in_T, out_T, W_T, narrow, 2368) \
+    f(in_T, out_T, W_T, narrow, 2432) \
     f(in_T, out_T, W_T, narrow, 2560) \
     f(in_T, out_T, W_T, narrow, 2752) \
     f(in_T, out_T, W_T, narrow, 2816) \
@@ -32,8 +37,12 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 3328) \
     f(in_T, out_T, W_T, narrow, 3456) \
     f(in_T, out_T, W_T, narrow, 3584) \
+    f(in_T, out_T, W_T, narrow, 3712) \
     f(in_T, out_T, W_T, narrow, 4096) \
+    f(in_T, out_T, W_T, narrow, 4480) \
     f(in_T, out_T, W_T, narrow, 4608) \
+    f(in_T, out_T, W_T, narrow, 4736) \
+    f(in_T, out_T, W_T, narrow, 4864) \
     f(in_T, out_T, W_T, narrow, 5120) \
     f(in_T, out_T, W_T, narrow, 5504) \
     f(in_T, out_T, W_T, narrow, 5632) \
@@ -43,8 +52,11 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 6848) \
     f(in_T, out_T, W_T, narrow, 6912) \
     f(in_T, out_T, W_T, narrow, 7168) \
+    f(in_T, out_T, W_T, narrow, 7424) \
     f(in_T, out_T, W_T, narrow, 8192) \
+    f(in_T, out_T, W_T, narrow, 8960) \
     f(in_T, out_T, W_T, narrow, 9216) \
+    f(in_T, out_T, W_T, narrow, 9472) \
     f(in_T, out_T, W_T, narrow, 10240) \
     f(in_T, out_T, W_T, narrow, 11008) \
     f(in_T, out_T, W_T, narrow, 11264) \
@@ -52,8 +64,11 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 13696) \
     f(in_T, out_T, W_T, narrow, 13824) \
     f(in_T, out_T, W_T, narrow, 14336) \
+    f(in_T, out_T, W_T, narrow, 14784) \
+    f(in_T, out_T, W_T, narrow, 14848) \
     f(in_T, out_T, W_T, narrow, 15360) \
     f(in_T, out_T, W_T, narrow, 16384) \
+    f(in_T, out_T, W_T, narrow, 18944) \
     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, 22528) \
@@ -61,6 +76,8 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, narrow, 27392) \
     f(in_T, out_T, W_T, narrow, 27648) \
     f(in_T, out_T, W_T, narrow, 28672) \
+    f(in_T, out_T, W_T, narrow, 29568) \
+    f(in_T, out_T, W_T, narrow, 29696) \
     f(in_T, out_T, W_T, narrow, 32000) \
     f(in_T, out_T, W_T, narrow, 32256) \
     f(in_T, out_T, W_T, narrow, 32512) \
@@ -85,9 +102,9 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
 // Keep above in sync with vllm/lora/layers::LogitsProcessorWithLoRA
 // and vllm/tests/lora/test_punica.py
 
-// Used for defining kernels going from the variety of 
+// Used for defining kernels going from the variety of
 // dim in to the narrow dim out
-    // Using it for the fully sharded column 
+    // Using it for the fully sharded column
     // parallel LoRA A which splits the rank dim
 #define FOR_INST_BGMV_NARROW(f, in_T, out_T, W_T, narrow) \
     f(in_T, out_T, W_T, 128, narrow) \
@@ -95,15 +112,20 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 512, narrow) \
     f(in_T, out_T, W_T, 640, narrow) \
     f(in_T, out_T, W_T, 768, narrow) \
+    f(in_T, out_T, W_T, 896, narrow) \
     f(in_T, out_T, W_T, 1024, narrow) \
     f(in_T, out_T, W_T, 1152, narrow) \
+    f(in_T, out_T, W_T, 1216, narrow) \
     f(in_T, out_T, W_T, 1280, narrow) \
     f(in_T, out_T, W_T, 1536, narrow) \
     f(in_T, out_T, W_T, 1664, narrow) \
     f(in_T, out_T, W_T, 1728, narrow) \
     f(in_T, out_T, W_T, 1792, narrow) \
     f(in_T, out_T, W_T, 2048, narrow) \
+    f(in_T, out_T, W_T, 2240, narrow) \
     f(in_T, out_T, W_T, 2304, narrow) \
+    f(in_T, out_T, W_T, 2368, narrow) \
+    f(in_T, out_T, W_T, 2432, narrow) \
     f(in_T, out_T, W_T, 2560, narrow) \
     f(in_T, out_T, W_T, 2752, narrow) \
     f(in_T, out_T, W_T, 2816, narrow) \
@@ -111,8 +133,12 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 3328, narrow) \
     f(in_T, out_T, W_T, 3456, narrow) \
     f(in_T, out_T, W_T, 3584, narrow) \
+    f(in_T, out_T, W_T, 3712, narrow) \
     f(in_T, out_T, W_T, 4096, narrow) \
+    f(in_T, out_T, W_T, 4480, narrow) \
     f(in_T, out_T, W_T, 4608, narrow) \
+    f(in_T, out_T, W_T, 4736, narrow) \
+    f(in_T, out_T, W_T, 4864, narrow) \
     f(in_T, out_T, W_T, 5120, narrow) \
     f(in_T, out_T, W_T, 5504, narrow) \
     f(in_T, out_T, W_T, 5632, narrow) \
@@ -122,8 +148,11 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 6848, narrow) \
     f(in_T, out_T, W_T, 6912, narrow) \
     f(in_T, out_T, W_T, 7168, narrow) \
+    f(in_T, out_T, W_T, 7424, narrow) \
     f(in_T, out_T, W_T, 8192, narrow) \
+    f(in_T, out_T, W_T, 8960, narrow) \
     f(in_T, out_T, W_T, 9216, narrow) \
+    f(in_T, out_T, W_T, 9472, narrow) \
     f(in_T, out_T, W_T, 10240, narrow) \
     f(in_T, out_T, W_T, 11008, narrow) \
     f(in_T, out_T, W_T, 11264, narrow) \
@@ -131,8 +160,11 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 13696, narrow) \
     f(in_T, out_T, W_T, 13824, narrow) \
     f(in_T, out_T, W_T, 14336, narrow) \
+    f(in_T, out_T, W_T, 14784, narrow) \
+    f(in_T, out_T, W_T, 14848, narrow) \
     f(in_T, out_T, W_T, 15360, narrow) \
     f(in_T, out_T, W_T, 16384, narrow) \
+    f(in_T, out_T, W_T, 18944, narrow) \
     f(in_T, out_T, W_T, 20480, narrow) \
     f(in_T, out_T, W_T, 22016, narrow) \
     f(in_T, out_T, W_T, 22528, narrow) \
@@ -140,6 +172,8 @@ void bgmv_kernel(out_T *__restrict__ Y, const in_T *__restrict__ X,
     f(in_T, out_T, W_T, 27392, narrow) \
     f(in_T, out_T, W_T, 27648, narrow) \
     f(in_T, out_T, W_T, 28672, narrow) \
+    f(in_T, out_T, W_T, 29568, narrow) \
+    f(in_T, out_T, W_T, 29696, narrow) \
     f(in_T, out_T, W_T, 32000, narrow) \
     f(in_T, out_T, W_T, 32256, narrow) \
     f(in_T, out_T, W_T, 32512, narrow) \
diff --git a/tests/lora/test_punica.py b/tests/lora/test_punica.py
index dae1d56874628..110c9b243507d 100644
--- a/tests/lora/test_punica.py
+++ b/tests/lora/test_punica.py
@@ -49,21 +49,30 @@ def _lora_ref_impl(
     128,
     256,
     512,
+    896,
     1024,
     1152,
+    1216,
     1280,
     1536,
     1664,
     2048,
+    2240,
     2304,
+    2368,
+    2432,
     2560,
     2752,
     3072,
     3328,
     3456,
     3584,
+    3712,
     4096,
+    4480,
     4608,
+    4736,
+    4864,
     5120,
     5504,
     5632,
@@ -73,19 +82,27 @@ def _lora_ref_impl(
     6848,
     6912,
     7168,
+    7424,
     8192,
+    8960,
     9216,
+    9472,
     10240,
     11008,
     11264,
     13824,
     14336,
+    14784,
+    14848,
     15360,
+    18944,
     22016,
     22528,
     24576,
     27392,
     27648,
+    29568,
+    29696,
     32000,
     32256,
     32512,