Guard blocksize=64 quantize instantiations for warp size compatibility

On AMD CDNA GPUs (warp size 64), blocksize=64 would mean only 1 thread
per warp in the quantize kernels, which is incompatible. Wrap these
instantiations with #if BNB_WARP_SIZE == 32 so they only compile on
NVIDIA.

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

Author: TimDettmers

csrc/kernels.cu

@@ -2461,36 +2461,46 @@ MAKE_kQuantizeBlockwise(half, 1024, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(half, 512, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(half, 256, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(half, 128, 2, 0, General8bit)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(half, 64, 2, 0, General8bit)
+#endif
 MAKE_kQuantizeBlockwise(half, 4096, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(half, 2048, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(half, 1024, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(half, 512, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(half, 256, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(half, 128, 2, 0, FP4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(half, 64, 2, 0, FP4)
+#endif
 MAKE_kQuantizeBlockwise(half, 4096, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(half, 2048, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(half, 1024, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(half, 512, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(half, 256, 2, 0, NF4)
 MAKE_kQuantizeBlockwise(half, 128, 2, 0, NF4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(half, 64, 2, 0, NF4)
+#endif
 MAKE_kQuantizeBlockwise(float, 4096, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(float, 4096, 4, 1, General8bit)
 MAKE_kQuantizeBlockwise(float, 2048, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(float, 1024, 4, 0, General8bit)
 MAKE_kQuantizeBlockwise(float, 512, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(float, 256, 2, 0, General8bit)
 MAKE_kQuantizeBlockwise(float, 128, 2, 0, General8bit)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(float, 64, 2, 0, General8bit)
+#endif
 MAKE_kQuantizeBlockwise(float, 4096, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(float, 2048, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(float, 1024, 4, 0, FP4)
 MAKE_kQuantizeBlockwise(float, 512, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(float, 256, 2, 0, FP4)
 MAKE_kQuantizeBlockwise(float, 128, 2, 0, FP4)
+#if BNB_WARP_SIZE == 32
 MAKE_kQuantizeBlockwise(float, 64, 2, 0, FP4)
+#endif
 MAKE_kQuantizeBlockwise(float, 4096, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(float, 2048, 4, 0, NF4)
 MAKE_kQuantizeBlockwise(float, 1024, 4, 0, NF4)