clean up some comments

Signed-off-by: jasonlizhengjian <[email protected]>

Author: jasonlizhengjian

vllm/compilation/fp8_allgather_pass.py

@@ -132,7 +132,7 @@ def __init__(self, config: VllmConfig):
         self.tp_size = get_tensor_model_parallel_world_size()
         if self.tp_size <= 1:
             self.disabled = True
-            logger.info(
+            logger.debug(
                 "FP8 AllGather optimization disabled: TP size = %d "
                 "(no communication needed)", self.tp_size)
             return
@@ -150,12 +150,12 @@ def __init__(self, config: VllmConfig):
                 self.tp_size,
                 self.tp_group_name,
             ).register(self.patterns)
-            logger.info(
+            logger.debug(
                 "FP8 AllGather optimization enabled: "
                 "TP size = %d, dtype = %s", self.tp_size, self.model_dtype)
         else:
             self.disabled = True
-            logger.info(
+            logger.debug(
                 "FP8 AllGather optimization disabled: "
                 "model dtype = %s (requires BF16)", self.model_dtype)
 
@@ -169,7 +169,7 @@ def __call__(self, graph: fx.Graph):
 
         self.matched_count = self.patterns.apply(graph)
         if self.matched_count > 0:
-            logger.info(
+            logger.debug(
                 "FP8 AllGather optimization: replaced %d AllGather "
                 "operation(s) with FP8 quantized versions", self.matched_count)
         else:

vllm/compilation/pass_manager.py

@@ -92,8 +92,6 @@ def configure(self, config: VllmConfig):
 
         if self.pass_config.enable_sequence_parallelism:
             self.passes += [SequenceParallelismPass(config)]
-            # FP8AllGatherOptPass must run BEFORE AsyncTPPass so that
-            # AsyncTPPass can fuse vllm_all_gather_fp8 + ScaledMM
             if self.pass_config.enable_fp8_allgather_opt:
                 self.passes += [FP8AllGatherOptPass(config)]
             if self.pass_config.enable_async_tp: