Repository navigation
Conversation
NVIDIA#6411 made qwen3-coder-tool declare <tool_call> an implicit end of reasoning, so a tool call emitted before </think> is split out of the reasoning text and parsed. That matches vLLM's combined qwen3 parser but not its standalone reasoning parsers paired with the qwen3_coder tool parser, which leave such a tool call in reasoning_content. Serving the same model on both engines with a separate reasoning parser now yields different tool calls. Keep qwen3-coder-tool strict and add qwen3-coder-tool-combined carrying the marker, so the behaviour is chosen in the parser list. Streaming markers are unchanged on both. Signed-off-by: Robert Kirby <rkirby@nvidia.com>
|
Auto-sync is disabled for draft pull requests in this repository. Workflows must be run manually. Contributors can view more details about this message here. |
sidsingh-nvidia
marked this pull request as ready for review
September 17, 2026 00:17
Contributor
|
/ok to test 9a6e2c7 |
kvareddy
approved these changes
Sep 17, 2026
sidsingh-nvidia
pushed a commit
to sidsingh-nvidia/Megatron-LM
that referenced
this pull request
Sep 17, 2026
PR NVIDIA#6411 made qwen3-coder-tool declare <tool_call> an implicit end of reasoning, so a tool call emitted before </think> is split out of the reasoning text and parsed. That matches vLLM's combined qwen3 parser but not its standalone reasoning parsers paired with the qwen3_coder tool parser, which leave such a tool call in reasoning_content. Serving the same model on both engines with a separate reasoning parser now yields different tool calls. Keep qwen3-coder-tool strict and add qwen3-coder-tool-combined carrying the marker, so the behaviour is chosen in the parser list. Streaming markers are unchanged on both. Folded into the Qwen3-Coder parser-parity branch alongside the grammar and coercion parity fixes (NVIDIA#7320); relevant to NVIDIA#7313. Signed-off-by: Robert Kirby <rkirby@nvidia.com> (cherry picked from commit c09b3fe) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com>
sidsingh-nvidia
pushed a commit
to sidsingh-nvidia/Megatron-LM
that referenced
this pull request
Sep 18, 2026
PR NVIDIA#6411 made qwen3-coder-tool declare <tool_call> an implicit end of reasoning, so a tool call emitted before </think> is split out of the reasoning text and parsed. That matches vLLM's combined qwen3 parser but not its standalone reasoning parsers paired with the qwen3_coder tool parser, which leave such a tool call in reasoning_content. Serving the same model on both engines with a separate reasoning parser now yields different tool calls. Keep qwen3-coder-tool strict and add qwen3-coder-tool-combined carrying the marker, so the behaviour is chosen in the parser list. Streaming markers are unchanged on both. Folded into the Qwen3-Coder parser-parity branch alongside the grammar and coercion parity fixes (NVIDIA#7320); relevant to NVIDIA#7313. Signed-off-by: Robert Kirby <rkirby@nvidia.com> (cherry picked from commit c09b3fe) Signed-off-by: Siddharth Singh <sidsingh@nvidia.com>
Contributor
|
Merged as a part of #7320. Closing. |
This branch was previously deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR do?
#6411 made
qwen3-coder-tooldeclare<tool_call>an implicit end of reasoning, so a tool call emitted before</think>is split out of the reasoning text and parsed. That matches vLLM's combinedqwen3parser engine, but not vLLM's standalone reasoning parsers (deepseek_r1and subclasses) paired with theqwen3_codertool parser, which leave such a tool call insidereasoning_content. Serving the same model on both engines with a separate reasoning parser now yields different tool calls, and RL rewards that penalise an unparsed<tool_call>see the two engines disagree.This keeps
qwen3-coder-toolstrict, matching the separate-parser pairing, and addsqwen3-coder-tool-combinedcarrying the marker for callers who want the combined-engine behaviour. The choice is made in the parser list, where it is visible in the server config, rather than implied by which tool parser is loaded. Streaming markers are unchanged on both.Issue tracking
Linked issue: none. Follow-up to #6411.
Contribution process
Pre-checks