FP4 Triton kernel bug fix (#4181)

Summary:
Pull Request resolved: #4181

X-link: facebookresearch/FBGEMM#1259

Fix loop iteration index calculation bug in triton kernel

Reviewed By: q10, jiawenliu64, jianyuh

Differential Revision: D75269590

fbshipit-source-id: c40bfeef4c487d9cdc0f0b1ba58071027a822961

Author: Tianyu Liang

fbgemm_gpu/experimental/gemm/triton_gemm/fp4_quantize.py

@@ -290,7 +290,7 @@ def _kernel_quantize_mx4_unpack(
         # Update offsets so we work on the next block.
         input_offset += GROUP_LOAD * GROUP_SIZE
         exp_offset += GROUP_LOAD
-        output_offset += GROUP_LOAD * GROUP_SIZE
+        output_offset += GROUP_LOAD * GROUP_SIZE // 2
 
 
 def triton_quantize_mx4_unpack(