feat(model-ports): add lfm25_8b_a1b, plamo_13b, ling_mini_2, phi35_moe - #214
Open
DarthCeltic wants to merge 5 commits into
Open
feat(model-ports): add lfm25_8b_a1b, plamo_13b, ling_mini_2, phi35_moe#214DarthCeltic wants to merge 5 commits into
DarthCeltic wants to merge 5 commits into
Conversation
Full compliant shape (claim + framework registration + real host-CPU perplexity). arch=lfm2moe, distinct from lfm2 (seed identity) -- Liquid AI's sparse-MoE variant. Built a plain CPU-only (GGML_ET=OFF) llama-perplexity from the vendored source, ran against the board- pinned WikiText-2 corpus. Loads and scores cleanly (PPL 29.8771), allocating both a transformer KV cache and a recurrent conv-state cache -- a genuine hybrid conv+attention+MoE architecture, confirmed live. First download attempt was truncated by a disk-space exhaustion issue on this session's working machine (unrelated to the model or repo) -- fixed by freeing ~79GB of already-verified, already-committed GGUFs from earlier in this campaign, then re-downloaded and re-verified complete and correct.
plamo_13b (arch=plamo, distinct from plamo2/plamo3) and ling_mini_2 (arch=bailingmoe2): full compliant shape, real host-CPU perplexity. Both loaded and scored cleanly after re-downloading past the same disk-space exhaustion issue documented on lfm25_8b_a1b (plamo_13b: 31.5102, ling_mini_2: 12.5899 -- third distinct MoE architecture this campaign to confirm MUL_MAT_ID works on ggml-cpu, after granite_3_1b_a400m and lfm25_8b_a1b). phi35_moe (arch=phimoe): its Q8_0 GGUF is 44.5 GB, the largest file attempted this campaign. Rather than commit most of the remaining disk budget to one file, only the reliable GGUF-header architecture read was kept (headers precede tensor data, so this is valid even from an intentionally-incomplete download) -- no host-CPU verification performed, disclosed honestly as the weakest-verified claim in this campaign rather than a full pass.
github-actions
Bot
added
track: model-ports
…ut negative results Both models load and run without crashing, but produce clearly degenerate output -- perplexity 5444 (hunyuan_0_5b) and 277 (exaone4_1_2b) against a normal range of ~5-70 for every working model in this campaign. Both size-verified against real remote Content-Length (not truncation). exaone4_1_2b uses the model's own official first-party GGUF, making a bad third-party quantization an unlikely explanation there specifically. Different failure category from bitnet_2b (fails to load) and persimmon_8b (architecture not compiled in) -- these load and run, but produce wrong output, likely a genuine graph-builder issue in this repo's vendored llama.cpp-et for these two architectures specifically. No claims filed for either.
…1b, ling_mini_2, plamo_13b Each claim's benchmark_config must exactly match the model's "config" pointer in .github/ci/benchmark_config.json per effective_model_config() in model_port_claim.py, not the top-level benchmark_config.json path itself. Fixes a mismatch that would fail eligibility validation.
Member
|
Manual ET review found mixed results, so this needs to be split/reworked:
Please keep only independently valid claims in each PR. |
…ms, pin revisions, add oracles Hackathon deadline has passed. This commit is record-completeness cleanup only, not a hackathon submission -- done so this catalog entry is accurate for ET-SoC1 card review. Per maintainer review: removed the lfm25_8b_a1b and phi35_moe model-port claims (LFM2.5-8B-A1B aborts on the ET CPY op; Phi-3.5-MoE's full Q8 exceeds board memory for full offload -- a partial-offload contract would be a different claim). Kept plamo_13b and ling_mini_2, which the maintainer's own manual ET-SoC1 review already reported passing: pinned their previously-floating 'main' HF revisions to exact commits, and added committed deterministic PPL oracles.
Contributor
Author
|
Note: this update was made after the official hackathon deadline (per maintainer's own announcement in the project Discord). It is not a hackathon submission and is not seeking credit or track scoring. We're cleaning up the remaining open PRs so this catalog is accurate for aiNekko's ET-SoC1 card records: removed the |
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
2 participants
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.
Adds lfm25_8b_a1b, plamo_13b, and ling_mini_2 causal-LM ports to llama.cpp-et with model-ports claims (real host-CPU perplexity for each). phi35_moe included with metadata-only verification (44.5GB file, disclosed honestly in its recipe rather than claiming a full pass).