Add MiniMax-M2, Mistral4, and GLM-DSA GGUF graphs - #686
Merged
Conversation
Performance Comparison
|
![github-code-quality[bot]](https://avatars.githubusercontent.com/u/9919?s=60&v=4){kind=small}
🏗️ Architecture Diff
No architecture changes detected. ✅ Legend: ⚪ No change · 🔵 Minor (attrs/inits) · 🟡 Moderate (nodes added/removed) · 🔴 Major (interface changed) |
Contributor
There was a problem hiding this comment.
Pull request overview
This PR adds explicit-float GGUF graph support for three previously deferred architectures (MiniMax-M2, Mistral4, GLM-DSA) by introducing new model/task wiring, GGUF tensor mapping + preprocessing, strict tensor-closure validation, and synthetic execution/evidence scaffolding. It promotes these routes in the GGUF architecture registry/census while keeping runtime support and quantized import fail-closed/deferred per the artifact budget constraints.
Changes:
- Add new explicit-float GGUF model/task implementations for MiniMax-M2 and Mistral4, plus complete GLM-DSA tensor routing and KV-B split handling.
- Enforce exact GGUF tensor closure for the promoted “remaining dense” cohort and add synthetic parity + metadata-only artifact evidence tests.
- Update GGUF architecture registry/census, loader pin inventory, and generated docs to reflect the newly importable routes and their deferred runtime/quantization boundaries.
Reviewed changes
Copilot reviewed 32 out of 32 changed files in this pull request and generated 1 comment.
Show a summary per file
| File | Description |
|---|---|
| tests/synthetic_parity_test.py | Excludes new GGUF-only model_types from HF-reference parity set. |
| tests/model_coverage_test.py | Documents new registered GGUF-only routes and their evidence boundaries. |
| tests/gguf_remaining_dense_artifact_test.py | Adds integration metadata-only Hub/LFS identity assertions for oversized candidates. |
| tests/build_graph_test.py | Marks mistral4_gguf as specialized (covered by dedicated GGUF tests). |
| tests/_test_configs.py | Adds tiny config for minimax_m2_gguf route. |
| src/mobius/tasks/_mistral4_gguf.py | New dedicated task with latent K-only cache ABI wiring for Mistral4 GGUF. |
| src/mobius/tasks/_glm_moe_dsa.py | Rejects static-cache for GLM-DSA and adds cache ABI/runtime-support metadata. |
| src/mobius/tasks/init.py | Exports/registers the new Mistral4 GGUF task. |
| src/mobius/models/glm_moe_dsa.py | Splits KV-B into K/V projections; adds weight preprocessing for fused KV-B input. |
| src/mobius/models/glm_moe_dsa_test.py | Strengthens preprocess-weights coverage (graph parameter coverage + KV-B split correctness). |
| src/mobius/models/gguf_mistral4.py | New explicit-float Mistral4 decoder with per-layer latent K-only cache. |
| src/mobius/models/gguf_minimax_m2.py | New explicit-float MiniMax-M2 decoder + sigmoid routing gate with correction bias. |
| src/mobius/models/init.py | Exports the new GGUF model classes. |
| src/mobius/integrations/gguf/_upstream_data/llamacpp_pin.json | Corrects Mistral4 loader census and tensor inventory to DeepSeek2 pin. |
| src/mobius/integrations/gguf/_tensor_processors.py | Adds split-MLA KV-B reshape/fuse processor; registers for GLM-DSA + Mistral4. |
| src/mobius/integrations/gguf/_tensor_processors_test.py | Adds GLM-DSA processor unit test for KV-B reshape. |
| src/mobius/integrations/gguf/_tensor_mapping.py | Adds GGUF→HF name maps for MiniMax-M2, Mistral4, and GLM-DSA. |
| src/mobius/integrations/gguf/_route_census_test.py | Updates expected category counts and asserts new evidence-only boundaries. |
| src/mobius/integrations/gguf/_remaining_dense.py | Adds strict closure validators for the three promoted routes. |
| src/mobius/integrations/gguf/_remaining_dense_test.py | Adds synthetic GGUF generation, closure tests, and executable prefill/decode parity. |
| src/mobius/integrations/gguf/_glm_moe_dsa_test.py | Updates GLM-DSA config/tensor mapping tests to official geometry + new routing. |
| src/mobius/integrations/gguf/_docs.py | Includes oversized-artifact evidence records in docs tables. |
| src/mobius/integrations/gguf/_conventional_decoder_test.py | Removes outdated Mistral4 conventional-decoder expectation. |
| src/mobius/integrations/gguf/_config_mapping.py | Adds key maps + postprocessors for MiniMax-M2/Mistral4/GLM-DSA and fail-closed rules. |
| src/mobius/integrations/gguf/_builder.py | Hooks remaining-dense tensor contract validation; adds static-cache/task restrictions. |
| src/mobius/integrations/gguf/_artifact_blocker_evidence.py | New immutable oversized-artifact evidence records + iteration API. |
| src/mobius/integrations/gguf/_arch_registry.py | Promotes MiniMax-M2/Mistral4 and completes GLM-DSA spec (tensor map + processor). |
| src/mobius/integrations/gguf/_arch_registry_test.py | Updates expected importable architecture counts and verdict coverage assertions. |
| src/mobius/components/_deepseek_mla.py | Adds optional split KV-B projection path for MLA component. |
| src/mobius/_registry.py | Registers minimax_m2_gguf and mistral4_gguf model types. |
| src/mobius/main.py | Adds CLI --static-cache rejections for GLM-DSA and Mistral4 GGUF. |
| docs/api/build_from_gguf.md | Regenerates GGUF support tables and adds the new artifact-budget blocker rows. |
💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.
justinchuby
force-pushed
the
justinchuby-add-dense-gguf-graphs
branch
from
August 28, 2026 06:01
958e485 to
d03ea6d
Compare
Implement exact explicit-float MiniMax-M2, Mistral4 latent-cache, and GLM-DSA import paths. Correct the pinned Mistral4 DeepSeek-V2 inventory, preserve packed/runtime fail-closed boundaries, and record immutable oversized artifact evidence. Co-authored-by: Copilot App <[email protected]> Copilot-Session: 0ee8718a-33b9-4015-9e20-3745bad54686 Signed-off-by: Justin Chu <[email protected]>
Canonicalize GLM-DSA alias dispatch while retaining raw metadata prefixes, force the dedicated task, and add alias-spelled weight/execution closure. Restore Mistral4 YaRN scaling from serialized metadata and add an independent mutation discriminator. Co-authored-by: Copilot App <[email protected]> Copilot-Session: 0ee8718a-33b9-4015-9e20-3745bad54686 Signed-off-by: Justin Chu <[email protected]>
Validate the serialized float32 yarn_log_multiplier before graph construction, accept the exact encoded [0.0, 0.1] endpoints, and reject non-finite or out-of-range values with real GGUF round-trip coverage. Co-authored-by: Copilot App <[email protected]> Copilot-Session: 0ee8718a-33b9-4015-9e20-3745bad54686 Signed-off-by: Justin Chu <[email protected]>
Pin the combined supported-architecture and remaining-route census after integrating the Grok, GroveMoE, and Hunyuan-MoE cohort from main with the MiniMax-M2, Mistral4, and GLM-DSA cohort. Co-authored-by: Copilot App <[email protected]> Copilot-Session: 0ee8718a-33b9-4015-9e20-3745bad54686 Signed-off-by: Justin Chu <[email protected]>
justinchuby
force-pushed
the
justinchuby-add-dense-gguf-graphs
branch
from
August 28, 2026 08:19
0e91cbb to
208f2cb
Compare
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.
Summary
This includes and supersedes the draft GLM-only work in #682.
Post-review fixes
glm_dsaaliases for validation/task/package dispatch while retaining the raw alias prefix for metadata reads; the alias-spelled GGUF proves the exact mixed full/shared indexer schedule, dedicated default task, complete indexer initializer materialization, and executable dense-MLA logitsrope.scaling.yarn_log_multiplierthroughyarn_apply_mscale; an independent metadata oracle pins0.0707, factor4, qk dim4to0.6028141, and mutation to0.1discriminates the result[0.0, 0.1]; the wire endpoint isfloat(np.float32(0.1)), used for both validation and normalization+inf,-inf, negative values, and the immediate float32 successor above0.1; realGGUFWriter/GGUFModelround trips accept both endpoints and prove finite attention scale and rotary caches before graph executionGlmMoeDsaTask()when callers omittaskEvidence and capability boundaries
Pinned semantic source:
ggml-org/llama.cpp@8d9af256337d1a501250f9bbf4c0859a654bddd6.All three real candidates exceed the 16 GiB bounded-artifact policy, so the PR records immutable metadata-only LFS evidence and uses reduced synthetic GGUFs for executable graph/cache parity:
minimax-m2mradermacher/MiniMax-M2-i1-GGUF@2d4f9b1a86d32ce4dfc47db312c8d6fcae8d7b37/MiniMax-M2.i1-IQ1_S.ggufmistral4unsloth/Mistral-Small-4-119B-2603-GGUF@bd93c721735aa32c035c0f19e738cb3371fd56ff/Mistral-Small-4-119B-2603-UD-IQ1_M.ggufglm-dsaunsloth/GLM-5.2-GGUF@abc55e72527792c6e77069c99b4cb7de16fa9f23/ sixUD-IQ1_SshardsNo GGUF payload was downloaded. Hub probes read file metadata only. Runtime support remains
DEFERRED, and quantization-preserving import isREJECTED; callers must usekeep_quantized=False. GLM routed DSA/MLA MTP, Mistral4 NextN/legacy MLA/Q-LoRA-free/nonzero-temperature layouts, and unowned GLM RoPE scaling all fail closed. There are no production model-ID or revision gates.Coverage waivers: real-weight L2/L4/L5 and runtime generation evidence cannot fit the artifact budget. Synthetic tests instead execute full-logit prefill/cached decode, MiniMax static cache, Mistral latent-vs-expanded MLA parity, GLM indexer scheduling, alias-spelled weight closure, and weight application.
Rebase reconciliation
Rebased onto #685 at
86e70a846b572a5266170420373781d2cac23450. The combined tree preserves the Grok/GroveMoE/Hunyuan-MoE and MiniMax-M2/Mistral4/GLM-DSA cohorts in registry dispatch, tensor-contract validation, architecture fingerprint isolation, model coverage, parity exclusions, census, and generated docs. Combined pins are 105 supported architectures and107dependency-blocked /100evidence-only /47immediately implementable remaining routes.Validation
2417 passed, 273 skipped3 passed8981 passed, 64 skipped, 12 deselected, 1 subtest passedlintrunner f --output oneline --all-filesandlintrunner -a --output onelinepython scripts/generate_gguf_support_docs.py --checkgit diff --check origin/main...HEAD0errors (29pre-existing production-file errors outside changed lines)Base:
86e70a846b572a5266170420373781d2cac23450Head:
208f2cb3c7c5215ea9bcf8547a54109acf768bee