[micahtyong]

Summary

Releases 0.1.4. This is the version to install for faster DeepSeek-V4.1-Flash sampling and sampler starts, and for spindle skills install.

• spindle.__version__ is now 0.1.4.
• The Unreleased section of CHANGELOG.md becomes 0.1.4 - 2026-10-29. It gains entries for the PRs merged after the v0.1.3 tag that had none: [trainer throughput] Train the DeepSeek-V4.1-Flash 32K recipe at TP4, and at TP2 on B300 #163 (32K at TP4, the deepseek_v41_flash_lora_32k_b300 recipe), [sampler throughput] Decode long DeepSeek-V4.1-Flash samples with the checkpoint's draft heads #165 (draft-head decoding on the 32K recipes), [sampler throughput] Let DeepSeek-V4.1-Flash samplers run 128 requests at once #147 (128 running requests), [sampler cold start] Load DeepSeek-V4.1-Flash sampler weights in parallel #148 and [sampler cold start] Pin Engram host tables per rank in DeepSeek-V4.1-Flash samplers #149 (sampler weight load and Engram pinning), [sampler cold start] Keep sampler kernels on the shared kernel-cache volume #146 (prebuilt prompt-indexing extension) and add skills and cli cmd #140 (spindle skills install). The two existing entries gain their PR numbers.
• The README install pins @v0.1.4.

When this merges, the release workflow tags v0.1.4 and publishes a GitHub Release from the 0.1.4 changelog section.

E2E

Checked from the release commit (f9976fc) before the merge: a fresh project with modal-spindle pinned to it, spindle deploy of a qwen35_9b_instruct_lora_16k subclass, then the stock Tinker RL example for 5 steps. spindle --version and /api/v1/get_server_capabilities both report 0.1.4 at that commit.

W&B: https://forge.coreweave.com/wandb/modal-labs/spindle-v014-e2e/runs/splns55o

Step	Reward	Step time
0	0.75	311.9 s (trainer and sampler cold start)
1	1.00	21.1 s
2	0.50	8.5 s
3	1.00	8.7 s
4	0.75	11.7 s

The 0.1.2 check of the same recipe took 319.5 s for step 0 and 18.5 to 26.2 s per step after.

🤖 Generated with Claude Code