chore: request handler fix for sglang + deepseekv4 end to end + sglang chat processor fix for deepseek v4 tool calling

[richardhuo-nv]

Summary

Make SGLang-backed Dynamo run DeepSeek-V4 end-to-end. Two fixes:

• Decode handler crash on older SGLang builds: DecodeWorkerHandler passes return_routed_experts unconditionally to Engine.async_generate, which fails on SGLang builds that don't have the kwarg yet.
• DeepSeek-V4 prompt rendering: DS-V4 ships without an HF chat template, so the existing tokenizer.apply_chat_template path produces wrong prompts for tool calls and thinking mode. DS-V4 needs SGLang's custom encoding_dsv4.encode_messages.

Changes

components/src/dynamo/sglang/_compat.py

• ensure_sglang_top_level_exports() — restores sgl.Engine / sgl.ServerArgs on SGLang 0.5.x snapshots that don't re-export them.
• filter_supported_async_generate_kwargs(engine, kwargs) — inspects Engine.async_generate signature (cached) and drops kwargs the installed engine can't accept; **kwargs signatures pass through unchanged.

components/src/dynamo/sglang/request_handlers/llm/decode_handler.py

• Route return_routed_experts through the filter at both prefill and decode async_generate call sites.

components/src/dynamo/frontend/sglang_prepost.py

• Detect DeepSeek-V4 by model/parser name when the tokenizer has no chat_template, and render via sglang.srt.entrypoints.openai.encoding_dsv4.encode_messages.
• Thinking mode derives from chat_template_kwargs.thinking; reasoning_effort is clamped to {max, high, None}.
• Tool filtering (handles tool_choice="none" and named tool choice) extracted into _filter_template_tools so both paths share it.
• Version-safe wrapper for SGLang's get_json_schema_constraint / parser.get_structure_constraint, which gained parallel_tool_calls in a later release.

components/src/dynamo/sglang/publisher.py

• Add from __future__ import annotations.

Tests

• components/src/dynamo/frontend/tests/test_sglang_processor_unit.py (+316): DS-V4 detection, tool filtering, thinking / reasoning-effort, parallel-tool-calls wrapper.
• components/src/dynamo/sglang/tests/test_sglang_unit.py (+98): async_generate kwarg filter.

Notes for reviewers

• Base branch is codex/deepseek-v4-parsers, not main — confirm merge order so this doesn't land orphaned.
• DS-V4 path raises ValueError at request time if the SGLang build lacks encoding_dsv4. Runtime-visible, not at startup — acceptable for now, flag if we'd rather fail fast at worker init.
• DS-V4 detection triggers when tokenizer.chat_template is None AND any of model / tool-call-parser / reasoning-parser normalizes to contain deepseekv4.

Test plan

• New unit tests pass (test_sglang_processor_unit.py, test_sglang_unit.py).
• E2E DS-V4 checkpoint with tool calling: named tool_choice, tool_choice="none", and parallel calls.
• Run against an SGLang build with return_routed_experts support and one without — confirm no crash in either case.

Where should the reviewer start?

Related Issues: (use one of the action keywords Closes / Fixes / Resolves / Relates to)

• closes GitHub issue: #xxx

---