Guard all blocksize=64 quantize instantiations for warp size compat

The previous commit missed the float/NF4 and all bnb_bfloat16
blocksize=64 instantiations. These use BLOCK_LOAD_WARP_TRANSPOSE
with 32 threads (64/2), which requires block_dim >= warp_size.
On CDNA (warp=64), 32 threads is insufficient.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: TimDettmers

csrc/kernels.cu

@@ -2507,7 +2507,9 @@ MAKE_kQuantizeBlockwise(float, 1024, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(float, 512, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(float, 256, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(float, 128, 2, 0, NF4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(float, 64, 2, 0, NF4)
+#endif
 
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 4096, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 4096, 4, 1, General8bit)
@@ -2516,21 +2518,27 @@ MAKE_kQuantizeBlockwise(bnb_bfloat16, 1024, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 512, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 256, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 128, 2, 0, General8bit)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 64, 2, 0, General8bit)
+#endif
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 4096, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 2048, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 1024, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 512, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 256, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 128, 2, 0, FP4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 64, 2, 0, FP4)
+#endif
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 4096, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 2048, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 1024, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 512, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 256, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 128, 2, 0, NF4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(bnb_bfloat16, 64, 2, 0, NF4)
+#endif
 
 // Template instantiations for blocksize=32 specialized kernel (4-bit only)
 #define MAKE_kQuantizeBlockwiseSmall(dtype, data_type_name)                                                            \