Skip to content

[spindle version] Release 0.1.8 - #360

Open
Micah Yong (micahtyong) wants to merge 2 commits into
mainfrom
micah/release-0.1.8
Open

Micah Yong (micahtyong) wants to merge 2 commits into
mainfrom
micah/release-0.1.8

Conversation

@micahtyong

@micahtyong Micah Yong (micahtyong) commented Nov 17, 2026 •

Copy link
Copy Markdown
Collaborator

Summary

Releases 0.1.8. The headline is that the Qwen3.5 LoRA recipes now train the linear-attention layers, which are three of every four layers. The dense Qwen and Nemotron-3 Nano LoRA recipes move to Miles native adapters, so checkpoints they saved before 0.1.8 load only on their original trainer image. A new recipe trains Nemotron 3.5 Super VL 120B-A12B on text. The release also returns temperature-1 logprobs for sampled tokens, scores multi-target batches with mixed candidate counts in one pass, and rewrites the profiling guide around tuning a deployment config.

When this merges, the release workflow tags v0.1.8 and publishes a GitHub Release from the 0.1.8 changelog section.

E2E

Ran on 83cc4da9, this PR rebased onto #365. The head is now bd9752d7, which adds only #361. #361 removes images that none of the configs below use. Clients used the stock Tinker SDK 0.39.1 through examples/shared-server/with_proxy_auth.py. get_server_capabilities reported 0.1.8 and 83cc4da9 on both deployments.

Qwen3.5-4B. A qwen35_4b_lora_32k subclass ran with two B200 trainer GPUs, one sampler replica and no lm_head target, because the stock RL example passes train_unembed=False. Images and adapter layout are the recipe's own. It was deployed with SPINDLE_TORCH_PROFILE_STEP=2.

Check Result W&B
Stock scripts/wandb_rl_example.py, 5 steps Step 0 159.5 s (cold start), then 8.0, 6.1, 7.5, 6.3 s. Loss and reward finite. hqk1mwsm
Profiled step 2 rank0.trace.json.gz and rank0.key_averages.txt written. Step 3's forward_backward took 0.61 s, against 0.56 s for step 2. same
Sampled logprobs at T=0.6 Mean |Δ| 0.007 vs compute_logprobs and 0.006 vs trainer forward (492 tokens) m89ucwt3
Mixed candidate counts, target_tokens [47, 3], [47, 1], [47, 5] Finite loss, output shapes match same
seed=1234 twice, one step each Max |Δ| 0.0 on trainer forward and compute_logprobs; 0.37 vs seed=4321 same
Exported adapter in_proj_q/k/v/z/a/b and out_proj adapters on all 24 linear-attention layers —

Nemotron 3.5 Super VL 120B-A12B. nemotron35-super-vl-120b-a12b-lora-64k ran as shipped: an 8×H200 trainer and a 4×H200 sampler, capped at one sampler replica. The recipe doesn't train lm_head, so the stock example's train_unembed=False matches. Text only.

Check Result W&B
Stock scripts/wandb_rl_example.py, 3 steps Step 0 870.5 s (cold start, including the weight download), then 84.3 and 69.5 s, mostly sampling. Loss and reward finite. b5qpbx95
Sampled logprobs at T=0.6 Mean |Δ| 0.030 vs compute_logprobs and 0.035 vs trainer forward (458 tokens) khkklyrh

spindle stop stopped both deployments afterwards.

Not run: the shared server, full fine-tuning and DFlash recipes.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

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