Repository navigation
bump(llama.cpp): server-cuda-b9967 → b10236 (DSpark; unblocks DeepSeek-Flash slugs) - #904
Merged
Merged
Conversation
Chosen because b10236 is the build the DeepSeek-V4-Flash offload matrix was measured on, and it carries DSpark (ggml-org#25784, "DeepseekV4 MTP + DSpark", merged 2026-08-02) which b9967 lacks — worth +26-56% decode on that model. This unblocks the DeepSeek-Flash slug work, whose entire evidence base sits on b10236 while the catalog shipped b9967. Deliberately NOT the newest build at bump time (b10297, same day): shipping an unmeasured image is the failure this pin exists to prevent (#187). DSpark presence verified in the images, not inferred: b9967 -> 0 `dspark` hits in libllama.so, --help does not advertise it b10236 -> 2 hits, `llama-server --help` advertises dspark Mechanical half via scripts/engine-pin-bump.sh; hand half per the pin-bump checklist: - engine profile spec + notes (pin history preserved) - 8 compose image defaults + the Tess compose header Engine: line - 2 pin-asserting test fixtures - docs/UPSTREAM.md engine-pin row - arch_patches.yml: N/A — the file tracks vllm/genesis only, no llama-cpp rows - vendored patches: N/A — `vendored_overlays: []` Validation: - llamacpp/mtp (Qwen3.6-27B single, draft-mtp): boot 16s, 3 warm-ups, verify-full ALL CHECKS PASSED. Running image confirmed by docker inspect. - llamacpp/tess-dual-mtp (dual, external draft-mtp): boot 16s, verify-full pass. - Tess streaming-tool-call caveat re-established: 12/12 on b10236 AND 12/12 on b9967 (targeted probe, matched payload). One isolated verify-full failure on b10236 was sampling variance -- that check sends temperature 0.3 with max_tokens 256 to a thinking model, so it is nondeterministic by construction. A single-run diff across two builds is not an A/B when the test is stochastic. - Full suite 98/98. Not run: quality-test.sh --medium (AGENTS.md pin-bump gate). The previous bump was think-OFF-neutral but +4 think-ON, so this is a real gap, deliberately deferred rather than skipped -- see PR discussion. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
…0236 The AGENTS.md pin-bump gate (quality-test.sh --medium) was not re-run. Maintainer decision 2026-08-06. Documented rather than silent because the compose Quality: lines carry pin provenance and would otherwise read as validation at the current pin. Functional validation IS current: verify-full on the qwen single and Tess dual paths, Tess streaming-tool-call caveat 12/12 on both pins, suite 98/98. The shipped quality numbers remain b9967-era. Stated as a live risk, not a formality: the previous bump (b9246->b9967) was think-OFF-neutral but +4 think-ON with 3 scenario flips, so a thinking-leg shift here would be invisible until someone runs the 8-pack. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
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.
Why now
b10236is the build the DeepSeek-V4-Flash offload matrix was measured on, and it carries DSpark (ggml-org#25784, "DeepseekV4 MTP + DSpark", merged 2026-08-02) — worth +26–56% decode on that model. The catalog shipsb9967, which lacks it, so the DeepSeek slug plan's entire evidence base sits on an image we don't ship. This is D2 of that plan and gates the rest of it.The pin is also ~26 days old, and
UPSTREAM.mdalready warns not to let it age.DSpark presence verified in the images, not inferred:
dsparkinlibllama.sollama-server --helpb9967b10236dspark✅b10297, published the same day). Shipping an unmeasured image is exactly the failure this pin exists to prevent (#187).What changed
Mechanical half via
scripts/engine-pin-bump.sh; hand half per the pin-bump checklist:spec:+notes:(pin history preserved)image:defaults, plus the Tess compose-headerEngine:linedocs/UPSTREAM.mdengine-pin rowarch_patches.yml: N/A — that file tracks vllm/genesis only; nollama-cpp-localrows existvendored_overlays: []Validation
llamacpp/mtp(Qwen 27B single, draft-mtp)llamacpp/tess-dual-mtp(dual, external draft-mtp)Running image confirmed by
docker inspect, not by the compose default.The one thing worth reading
A single
verify-fullrun onb10236failed check 6 (streaming tool-calls,finish=length) whileb9967passed — which looks exactly like a regression of the Tess "RESOLVED CAVEAT" that was declared closed on 3/3 clean.It isn't. Targeted probe, payload matched byte-for-byte:
b9967b10236That check sends
temperature 0.3withmax_tokens 256to a thinking model — the reasoning trace and the tool call must both fit, so it is nondeterministic by construction. A single-run diff across two builds is not an A/B when the test is stochastic. Reporting the first result would have blocked this bump over nothing. Net effect: the caveat's evidence is upgraded from 3/3 at one build to 12/12 at each of two.quality-test.sh --medium(theAGENTS.mdpin-bump gate) has not run. The previous bump (b9246→b9967, #680) was think-OFF-neutral but +4 think-ON with 3 scenario flips, so this is a genuine gap rather than a formality — flagged deliberately rather than skipped.