[Docs] Add comment explaining DEEPEP_MAX_TOKENS_PER_RANK env variable (#1624)

claude[bot] · github-actions[bot] · web-flow · commit 59120b682882 · 2026-03-27T10:05:04.000+08:00
Co-authored-by: claude[bot] &lt;41898282+claude[bot]@users.noreply.github.com&gt;
diff --git a/xtuner/v1/ray/rollout/lmdeploy.py b/xtuner/v1/ray/rollout/lmdeploy.py
@@ -349,6 +349,11 @@ def _transform_rollout_config_to_server_configs(self) -> Namespace:
                     {
                         "LMDEPLOY_DP_MASTER_ADDR": dist_addr,
                         "LMDEPLOY_DP_MASTER_PORT": dist_port,
+                        # DEEPEP_MAX_TOKENS_PER_RANK is required by DLBlas's DeepEP
+                        # token dispatcher used in lmdeploy EP mode. Without it,
+                        # lmdeploy will fail during warmup.
+                        # Ref: https://github.com/DeepLink-org/DLBlas/blob/aae23445/dlblas/layers/moe/token_dispatcher.py#L81
+                        # Ref: https://github.com/InternLM/lmdeploy/blob/81627e3d/lmdeploy/utils.py#L375
                         "DEEPEP_MAX_TOKENS_PER_RANK": str(max_batch_size),
                     }
                 )

Original file line number	Diff line number	Diff line change
`@@ -349,6 +349,11 @@ def _transform_rollout_config_to_server_configs(self) -> Namespace:`
`349`	`349`	`{`
`350`	`350`	`"LMDEPLOY_DP_MASTER_ADDR": dist_addr,`
`351`	`351`	`"LMDEPLOY_DP_MASTER_PORT": dist_port,`
	`352`	`+ # DEEPEP_MAX_TOKENS_PER_RANK is required by DLBlas's DeepEP`
	`353`	`+ # token dispatcher used in lmdeploy EP mode. Without it,`
	`354`	`+ # lmdeploy will fail during warmup.`
	`355`	`+ # Ref: https://github.com/DeepLink-org/DLBlas/blob/aae23445/dlblas/layers/moe/token_dispatcher.py#L81`
	`356`	`+ # Ref: https://github.com/InternLM/lmdeploy/blob/81627e3d/lmdeploy/utils.py#L375`
`352`	`357`	`"DEEPEP_MAX_TOKENS_PER_RANK": str(max_batch_size),`
`353`	`358`	`}`
`354`	`359`	`)`