Revert "[RL] Support Rollout Routing Replay"

[Jiang-Jia-Jun]

Reverts #5321

[paddle-bot]

Thanks for your contribution!

[Copilot]

Pull request overview

This PR reverts the "Rollout Routing Replay" feature (PR #5321), systematically removing all related code, configuration, and test changes. The revert cleanly removes the routing replay functionality that was added for RL training scenarios.

Key changes:

• Removed RoutingReplayConfig class and all routing replay configuration parameters from the configuration system
• Deleted the entire routing_indices_cache.py file containing RoutingReplayManager and routing store implementations
• Removed topk_ids_hookfunc callback parameter from all MoE backend implementations
• Cleaned up routing replay manager initialization and usage in gpu_model_runner.py

Reviewed changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
fastdeploy/config.py	Removed RoutingReplayConfig class and routing_replay_config parameter from FDConfig
fastdeploy/engine/args_utils.py	Removed routing_replay_config CLI argument and create_routing_repaly_config method
fastdeploy/engine/engine.py	Removed routing_replay_config from worker service startup arguments
fastdeploy/worker/worker_process.py	Removed RoutingReplayConfig import and routing_replay_config initialization
fastdeploy/worker/gpu_model_runner.py	Removed RoutingReplayManager initialization, is_chunk_step tracking, and routing replay logic
fastdeploy/model_executor/forward_meta.py	Removed routing_replay_table field from ForwardMeta
fastdeploy/model_executor/layers/moe/routing_indices_cache.py	Deleted entire file containing routing replay functionality
fastdeploy/model_executor/layers/moe/moe.py	Removed enable_routing_replay flag, topk_ids_hookfunc parameter, and routing replay logic from forward methods
fastdeploy/model_executor/layers/moe/fused_moe_*.py	Removed topk_ids_hookfunc parameter from all backend apply methods (cutlass, deepgemm, marlin, triton, wint2)
fastdeploy/model_executor/layers/backends/*	Removed topk_ids_hookfunc parameter from backend-specific MoE implementations (metax, gcu, dcu)
fastdeploy/model_executor/models/glm4_moe.py	Removed default parameter from Glm4Moe.forward signature
fastdeploy/rl/rollout_config.py	Removed routing_replay_config parameter
tests/layers/test_*.py	Removed RoutingReplayConfig imports and initialization from test files
tests/e2e/test_EB_Lite_serving.py	Removed --routing-replay-config command line arguments
tests/distributed/chunked_moe.py	Removed routing_replay_manager and enable_routing_replay from mock objects

[Copilot]

[nitpick] This formatting change (splitting the function call across multiple lines) is not necessary for the revert and adds noise to the diff. The original single-line format hidden_states = self.mlp(hidden_states, forward_meta) would be sufficient. Consider keeping the formatting consistent with the rest of the codebase unless there's a specific style guide reason for the change.

Suggested change

	hidden_states = self.mlp(
	hidden_states,
	forward_meta,
	)
	hidden_states = self.mlp(hidden_states, forward_meta)

[Copilot]

[nitpick] The forward_meta parameter lacks a type annotation. For consistency with the rest of the codebase (e.g., FusedMoE.forward in moe.py line 625 has forward_meta: ForwardMeta), consider adding the type hint: def forward(self, x, forward_meta: ForwardMeta):

Suggested change

	def forward(self, x, forward_meta):
	def forward(self, x, forward_meta: ForwardMeta):