Remove one kernel in per_tensor_quant_mla_fp8 (#5549)

python/sglang/srt/utils.py

@@ -1932,3 +1932,16 @@ def is_fa3_default_architecture(hf_config):
         "MistralForCausalLM",
     }
     return architectures[0] in default_archs
+
+
+# Can be more general if it is used in multiple places (keep it simple and thus not general now)
+class BumpAllocator:
+    def __init__(self, buffer_size: int, dtype, device):
+        self._buffer = torch.zeros((buffer_size,), dtype=dtype, device=device)
+        self._pointer = 0
+
+    def allocate(self, size: int):
+        assert self._pointer + size <= len(self._buffer)
+        output = self._buffer[self._pointer : self._pointer + size]
+        self._pointer += size
+        return output