Deepseek V4 #23882

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged

hnyls2002 merged 428 commits into main from dsv4-rebase

May 8, 2026

Merged

Deepseek V4 #23882

rebase main with d72f58d

Loading branch information

hnyls2002 committed Apr 29, 2026

commit c93d658fb9d558745c0198e9e115f0133b4e0c1d

python/sglang/srt/managers/scheduler.py

ispobock marked this conversation as resolved.

-Original file line number
+Diff line change
@@ Expand Up @@
                     else:
                         self.running_batch.merge_batch(new_batch)
                     self.running_batch.hisparse_coordinator = self.hisparse_coordinator
+                # Reset batch_is_full so the scheduler can schedule more prefills.
                 self.running_batch.batch_is_full = False
             if (
@@ Expand Down @@

python/sglang/srt/managers/scheduler_output_processor_mixin.py

-Original file line number
+Diff line change
@@ Expand Up / @@ -472,29 +472,7 @@ def process_batch_result_decode( @@
                 self._handle_finished_req(req, i, logits_output)
-                if req.finished():
-                    # delete feature to save memory
-                    if req.multimodal_inputs is not None and req.session is None:
-                        req.multimodal_inputs.release_features()
-                    self.maybe_collect_routed_experts(req)
-                    self.maybe_collect_indexer_topk(req)
-                    if self.server_args.disaggregation_decode_enable_offload_kvcache:
-                        # Asynchronously offload KV cache; release_kv_cache will be called after Device->Host transfer completes
-                        if not self.decode_offload_manager.offload_kv_cache(req):
-                            self.decode_offload_manager.finalize_release_on_finish(req)
-                    else:
-                        if self.enable_hisparse:
-                            self.hisparse_coordinator.request_finished(req)
-                        release_kv_cache(req, self.tree_cache)
-                    req.time_stats.set_completion_time()
-                self.maybe_collect_customized_info(i, req, logits_output)
-                if req.return_logprob and (
-                    batch.spec_algorithm.is_none() or batch.is_spec_v2
-                ):
+                if req.return_logprob:
                     # Spec v1 handles logprobs inside its own worker.
                     # Normalize: non-spec has 1 token, spec v2 has multiple.
                     if batch.is_spec_v2:
@@ Expand Down Expand Up / @@ -574,6 +552,7 @@ def _handle_finished_req( @@
                 if req.multimodal_inputs is not None and req.session is None:
                     req.multimodal_inputs.release_features()
                 self.maybe_collect_routed_experts(req)
+                self.maybe_collect_indexer_topk(req)
                 if self.server_args.disaggregation_decode_enable_offload_kvcache:
                     # Asynchronously offload KV cache; release_kv_cache will be called after Device->Host transfer completes
@@ Expand Down Expand Up / @@ -986,7 +965,6 @@ def stream_output_generation( @@
             retraction_counts = []
             output_hidden_states = None
             load = self.get_load()
-            indexer_topk = None
             routed_experts = None
             customized_info = {}
@@ Expand Down Expand Up / @@ -1171,10 +1149,7 @@ def stream_output_generation( @@
                         if routed_experts is None:
                             routed_experts = []
                         routed_experts.append(req.routed_experts)
-                    if req.return_indexer_topk:
-                        if indexer_topk is None:
-                            indexer_topk = []
-                        indexer_topk.append(req.indexer_topk)
                     if req.customized_info is not None:
                         for k, v in req.customized_info.items():
                             if k not in customized_info:
@@ Expand Down Expand Up / @@ -1230,7 +1205,6 @@ def stream_output_generation( @@
                         output_token_entropy_val=None,
                         output_hidden_states=output_hidden_states,
                         routed_experts=routed_experts,
-                        indexer_topk=indexer_topk,
                         customized_info=customized_info,
                         placeholder_tokens_idx=None,
                         placeholder_tokens_val=None,
@@ Expand Down @@

You are viewing a condensed version of this merge commit. You can view the full changes here.

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Deepseek V4 #23882

Uh oh!

Diff view

Diff view

Uh oh!

There are no files selected for viewing

Uh oh!

Uh oh!