fix(llamacpp): pin mainline build by digest, retire dead TurboQuant fork#54
Merged
Merged
Conversation
The llamacpp image was tagged as the official `server-cuda` but actually ran a TurboQuant fork build (`0a8062d`, ~April) — three disagreeing sources of truth: the `.env` override, the rolling tag, and `llamacpp/Dockerfile` still building the fork. That old build predates upstream MTP support (PR #22673), so Qwen3.6 MTP GGUFs failed to load (`missing tensor blk.N.ssm_conv1d.weight`). - Pin both llama.cpp services to the current mainline image BY DIGEST (`ghcr.io/ggml-org/llama.cpp@sha256:44cd0833…`, build 9765 / 73618f27a) — the first build that loads Qwen3.6 MTP + qwen35moe natively. - Remove the dead `build:` stanza and `llamacpp/Dockerfile` (the unused fork). - Reconcile `.env.example` to the single pinned source of truth. Validated end-to-end: recreated llamacpp on build 9765; the A3B (qwen35moe) loads on GPU at full 512K with the `qwen35moe.context_length` override + YaRN + mmproj and serves via model-gateway (system_fingerprint `b9765-73618f27a`). The dense Qwen3.6-27B MTP GGUF also loads natively on this build — no per-model tensor stripping. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The fork is gone (mainline pinned by digest), so .env.example must not advertise tbq*/tbqp* KV cache types — they don't exist on mainline and an operator setting one would crash llama-server. Replace with the mainline KV types (q8_0 default for the hybrid A3B) and drop the build-specific 0a8062d reference. Co-Authored-By: Claude Opus 4.8 (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
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.
Problem (config drift)
The
llamacppimage was tagged as officialserver-cudabut actually ran a TurboQuant fork build (0a8062d, ~April) — three disagreeing sources of truth: the.envoverride, the rolling tag, andllamacpp/Dockerfilestill building the fork. That old build predates upstream MTP support (PR #22673), so Qwen3.6 MTP GGUFs failed to load (missing tensor blk.N.ssm_conv1d.weight).The previous workaround was to strip the MTP head off every model — a per-artifact bandaid. The real fix is updating the engine.
Change
ghcr.io/ggml-org/llama.cpp@sha256:44cd0833…(build 9765 /73618f27a) — the first build that loads Qwen3.6 MTP +qwen35moenatively. No rolling tags.build:stanza +llamacpp/Dockerfile(the unused fork) → one source of truth..env.example.Validation (in production, before commit)
Recreated
llamacppon build 9765 via the blessed two-env-file mechanism. The A3B (qwen35moe) loads on the GPU at full 512K with theqwen35moe.context_lengthoverride + YaRN 2× + mmproj, and serves through model-gateway —system_fingerprint: b9765-73618f27a, RTX 5090 at 28.8/32.6 GB. The dense Qwen3.6-27B MTP GGUF also loads natively on this build (CPU-isolated test) — the per-model strip is retired.docker compose configparses clean with both services on the digest, no build stanza.Follow-ups (separate, not bundled)
local-chat→ dense Qwen3.6-27B (native 256K,--spec-type draft-mtp).stack_monitorto track the pinned llama.cpp build number (currently ROLLING).🤖 Generated with Claude Code