qwen3.8-27b: fold in Defilan #24 field notes (MTP, gfx1151, template fixes) - #25
Merged
Conversation
…corrections) Adds the MTP speculative-decoding section the card lacked (head ships as a separate 1.6 GiB GGUF in ggml-org, not the tested unsloth quant; up to ~3.5x, workload-dependent optimum matching our independent vLLM/SGLang runs), a Strix Halo gfx1151 throughput row with two Vulkan cautions, the reasoning_effort:high HTTP 500 correction, the medium-reasons-more-than-low addendum to Trap 2, and the second output-budget cause of the turn-6 signature. Closes #24 field notes; every existing finding Defilan probed reproduced.
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
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Folds the field notes from #24 (thanks @Defilan) into the Qwen3.8-27B guide. Every existing finding he probed reproduced; these are additions, not corrections to the behavioral results.
What landed
ggml-org/Qwen3.8-27B-GGUF, not in the tested unsloth quant, which is why the card had no MTP section. Up to ~3.5x single-stream, workload-dependent optimum (code keeps climbing to nDraftMax 7, prose peaks at 3, acceptance falls with depth). This matches our independent vLLM/SGLang runs on the same GB10 (DSpark/EAGLE 45-53 on structured, ~16 on prose).reasoning_effort: high= HTTP 500 on the stock template, not a quiet degrade.mediumis inert as an instruction but reasons more thanlow.max_tokens), distinct from context exhaustion, distinguishable by checkingcompletion_tokensagainst both bounds.check_guides passes locally (frontmatter + no em/en-dashes).