Blog: Running Qwen3.8-27B on DGX Spark - #262
Merged
Merged
Conversation
Deploying website with
|
| Latest commit: |
e6e1328
|
| Status: | ✅ Deploy successful! |
| Preview URL: | https://cf1bc8b1.website-dab.pages.dev |
| Branch Preview URL: | https://qwen38-spark.website-dab.pages.dev |
Preflight against the saved benchmark CSVs turned up a few things: - Vision section claimed the model read "11.3 t/s" off the chart; the chart shows 11.6 (llama.cpp Q4 single stream). Corrected. - Vision section referenced "the chart you saw above" but sat 30 lines before the chart. Moved it after the Ollama section so the callback works. - Wrapping-up context-flatness bullet mixed engines (8.2 and 7.9 are FP8, 9.8 is NVFP4). Now all NVFP4: 11.5 -> 10.9 -> 9.8. - 32K decode drop is 3%, not 4%. Requant delta is within 4%, not 3%, and only the NVFP4 revision has a saved re-run, so the GGUF claim is dropped. - Test environment paragraph is now a table (unreadable on mobile as prose), and it names the Aug 17 runs that produced the MTP/SGLang/DSpark numbers. - qwen3_5 architecture explanation appeared three times; the causal version now lands once and llama-bench keeps only the observation. - 121GB vs 128GB contradiction reworded; added the bandwidth-ceiling arithmetic so 11.6 t/s reads as ~70% of peak rather than "close to" it. - New light-theme cover reflecting the finished post (4 engines, 105.8 t/s peak aggregate) instead of the day-zero-only framing.
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
Day-zero to day-three coverage of Qwen3.8-27B on the DGX Spark (GB10): recipes and measured numbers for llama.cpp, Ollama, vLLM (FP8/NVFP4, with and without MTP), and SGLang (with and without NEXTN speculation).
Includes three sharp-edge findings: the FP8 deep-context concurrency wedge, the MTP hard-reboot at 16K+concurrency (reproduced 2/2), and a stale SGLang container silently generating garbage behind a green health check.
Numbers all trace to llama-benchy/llama-bench runs; Spark Arena submission sub1786754097881. Draft for preview/read-through.