Skip to content

[model support] Add Inkling-2-Small attention multi-LoRA - #342

Open
Kailash Ranganathan (kailash109) wants to merge 16 commits into
mainfrom
codex/inkling2-multilora
Open

Kailash Ranganathan (kailash109) wants to merge 16 commits into
mainfrom
codex/inkling2-multilora

Conversation

@kailash109

@kailash109 Kailash Ranganathan (kailash109) commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Add inkling2-small-lora-8k-b300 for two independent rank-32 attention and unembedding adapters on the full Inkling-2-Small checkpoint. Training uses one eight-B300 node; sampling uses a separate eight-B300 node. Stock Tinker clients set train_mlp=False; experts and attention sinks stay frozen.

The trainer pins Miles PR 45 at e186d97710dcc319da659f8bc25dbb674e227909 on spindle-miles-v0.2.0. A dedicated frozen sampler image selects Inkling precision metadata, preserves fused attention target paths, and adds the release image's missing unembedding target alias. Add opt-in inference RDMA allocation so full-node samplers honor reserved-capacity exclusion.

Register the recipe in the deployment picker and script, validate supported parallelism, and include reproducible SDK and image checks. Existing named images retain their contents.

Validation

Full-model validation ran Spindle 3e0139d56b726eef8c860e639c0518150dab6bd1 and Miles e186d977 on separate eight-B300 trainer and sampler nodes, with exclude_reserved_capacity=True:

  • Stock SDK 0.26.0: two adapters, forward/backward, Adam updates, publication, and actual SGLang sampling passed. A's mean loss fell from 2.8943 to 2.6545; B stayed exactly at 2.8943. A's serving logprobs changed by up to 0.866795 nats; B stayed unchanged within 0.00001 nats.
  • Stock tinker-cookbook 0.7.0 with SDK 0.39.1 restored the attention-only checkpoint, trained one step on 32 tokens (mean NLL 3.41838), and saved final training and sampler checkpoints. The existing helper supplies proxy authentication headers; the SDK and cookbook are unmodified.
  • Save, mutate, and restore weights and optimizer returned A's exact saved loss.
  • 8,192-token training and publication passed. Long-prompt sampling accepted 8,185 prompt tokens and returned five finite-logprob tokens.
  • Warm short backward: 2.62 seconds; optimizer step: 0.49 seconds. Warm 8K backward: 29.64 seconds. Earlier uncached runs took 327.61 and 716.61 seconds respectively, including compilation. These are observed call times, not performance comparisons.
  • Full CPU suite on Python 3.12.16: 1,572 passed, 11 skipped after the final main merge. All pre-commit checks passed. Minimum, image-pinned, and latest SDK CI passed on the validated runtime.
  • The newer merged Miles PR head d2336710 separately passed 251 native-LoRA tests and 10 Inkling launcher snapshots in the matched image on B300. Full-model results apply to frozen e186d977, not that newer head. The full launcher suites had the same 23 failures and four errors on head and unchanged e8142459 base; no new failing cases (679 versus 669 passing).

B200, colocation, image inputs, engine restart, convergence, and full-model mixed-client batches were not validated. Requests used two resident clients sequentially; separate CUDA tests cover mixed-slot forward and gradient parity. See docs/inkling2-validation.md and its sanitized JSON evidence for full results. After the GPU run, main through 0d8f611 was merged to resolve a changelog conflict. The later profiling, scoring, deployment typing, and usage changes have CPU coverage; they were not rerun on the full model.

Temporary validation apps were stopped; a container listing confirmed no remaining validation GPU containers. Shared model assets and checkpoints were retained.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant