fix: tmp revert gpt oss tp sharding on hopper (#9469)

python/sglang/srt/models/gpt_oss.py

@@ -793,9 +793,12 @@ class GptOssForCausalLM(nn.Module):
             intermediate_size % mxfp4_block == 0
         ), f"{intermediate_size=} must be divisible by {mxfp4_block=}"
         intermediate_size_block = intermediate_size // mxfp4_block
-        per_rank_intermediate_size_block = math.ceil(
-            intermediate_size_block / moe_tp_size
-        )
+        if _is_sm100_supported:
+            per_rank_intermediate_size_block = math.ceil(
+                intermediate_size_block / moe_tp_size
+            )
+        else:
+            per_rank_intermediate_size_block = intermediate_size_block // moe_tp_size
         per_rank_intermediate_size = per_rank_intermediate_size_block * mxfp4_block
 
         # Calculate common slicing bounds for current rank