Replace estimates in the Qwen3.8-27B profile with measurements; add launch script and notes - #233
Merged
Merged
Conversation
…aunch script and notes scripts/settings_local_llm_qwen38_27b.yaml said "UNVALIDATED", estimated ~65-75 tok/s plain for IQ3_XXS and 1.3x-2.8x from MTP, and set spec_draft_n_max to 2. Measured on an RTX 5070 Ti (16 GB, WSL2, llama.cpp 3173a56): 54.6 tok/s plain, 76.4 (n=3) / 78.2 (n=4) tok/s on 69 replayed real agent requests, 14,128 MiB peak VRAM. n=4 is the best of 2-6; a separate DFlash2 drafter ties it on real traffic for about 1 GB more VRAM. The profile comments now say what was measured and where (new docs/local_qwen38_27b.md, with the speed/VRAM table, prefix-cache behaviour, a small SWE-bench Lite dev screen with its caveats, and the known issues), and scripts/serve_qwen38_27b_llamacpp.sh starts the measured configuration.
🤖 Godspeed Review
SecurityNo secrets detected in changed files. |
🤖 Godspeed Review
SecurityNo secrets detected in changed files. |
🤖 Godspeed Review
SecurityNo secrets detected in changed files. |
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Why
scripts/settings_local_llm_qwen38_27b.yaml(merged in #228) said "UNVALIDATED", estimated ~65-75 tok/s plain for the IQ3_XXS file and 1.3x-2.8x from MTP, and setspec_draft_n_max: 2as "community-tuned". I have now measured it on the target machine, and the estimates were wrong (the plain-decode one by about 20-30%).Measured (RTX 5070 Ti 16 GB, WSL2, llama.cpp 3173a56, 32K context, q8_0 KV)
"Real agent requests" = 69 requests captured from an agent-in-loop run and replayed against each config; the live evaluation of MTP n=4 then measured 77.5 tok/s over 249 requests. Synthetic prompts overstate the speed-up because real output is mostly reasoning prose that the drafters accept poorly.
Changes
spec_draft_n_max: 4.docs/local_qwen38_27b.md: speed/VRAM table, prefix-cache behaviour on the real Godspeed path, a small SWE-bench Lite dev screen (27B: 19/24 over three draws; KAT REAP-50: 12/24 over three) with its caveats (oracle tool in the loop, 8 tasks, draws not independent, protocol differences between draws), and known issues.scripts/serve_qwen38_27b_llamacpp.sh: the measured launch command.Verification
bash -non the script;tests/test_config_llamacpp.pyandtests/test_tools/test_llamacpp_manager.pypass (102 passed, 1 skipped);ruff check .clean.Not verified
q4_0quality, other reasoning efforts, and the IQ3_S file were not compared on quality.compaction_threshold: 0.7) is a suggestion derived from the measured 1.29x estimator gap, not something I ran.