[listar2000]

Outcome

main takes the next upstream image, which adds upstream's native LoRA framework. This is the base the DeepSeek-V4.1 LoRA work builds on; on its own it changes nothing for the recipes we run.

	before	after
upstream base	af4413b74 (2026-10-14)	69a147dfc (2026-10-16), 6 upstream commits
base image	radixark/miles:dev b836fec…	radixark/miles:dev-202610160544 c64fa34…
SGLang / Megatron-LM in the image	7b6993a / 9df207f	same
native LoRA for raw mode (miles/backends/megatron_utils/lora/native/)	no	yes (upstream #1792)

69a147dfc is the first upstream commit after #1792 that a published upstream image was built from (read from the image's /root/miles/.git), as the rulebook requires for a between-releases sync.

Contract

• Base image changes to radixark/miles:dev@sha256:c64fa344e3a86715f03c0a1b6a422c474e0035355591a2f6da87bf52a74c52da; same CUDA, torch, SGLang and Megatron commits. Python packages are unchanged: no docker/, requirements.txt or pyproject.toml change in the range.
• W&B and image labels show upstream:v0.1.2+69a147dfc2.
• Upstream removed --log-probs-max-tokens-per-gpu (#4118). No recipe in miles_ext sets it.

Side-effects

The six upstream commits:

• #1792 native LoRA for raw mode. It replaces the Kimi K3 and Inkling LoRA plugins (miles_plugins/models/{kimi_k3,inkling}/lora.py are deleted) with a registry and per-architecture specs. Only active with --lora-* flags.
• #4099 torch-native per-token loss normalizer. Torch-native backend only; the Megatron loss path does not pass the new argument.
• #4114 custom modules with --skip-actor-forward-only. Opt-in.
• #4118 removes --log-probs-max-tokens-per-gpu.
• #4120 (test) and #4112 (docs).

No merge conflicts; no fork fix or hook needed re-applying. Open PRs #36 and #45 do not touch the changed files.

Done

• Merge of upstream 69a147dfc (Fork-Infra: sync-69a147dfc).
• miles_ext/stack.py: UPSTREAM_BASE and BASE_IMAGE for that commit.

Verification

• python fork/tools/ledger.py check: ok (119 fork commits, 43 upstream files touched).
• fork/vendored.json: the only vendored file (Harbor sail.py) is outside the range.
• REVIEW_PATHS diff (fork/tools/sync.py): only miles/utils/arguments.py (the flag removal above).
• CPU tests in the miles-mobius image with this tree: pytest miles_ext/tests tests/fast/backends/megatron_utils/lora tests/fast/backends/megatron_utils/test_{lora_utils,lora_model_branches,checkpoint,model_initialize}.py: 419 passed, 10 skipped.
• Base image pins, read from the new image's .git refs (2026-10-19): Miles 69a147d, SGLang (sglang-miles) 7b6993a31, Megatron-LM (miles-main) 9df207fd1. SGLang and Megatron are the same commits as the previous base image, so the BASE_IMAGE comment's "plus 9 SGLang and 1 Megatron commits" still holds.
• Flag-off GPU step on the new base image (4-layer DeepSeek-V4.1 slice, 1x B300:8, GPU check B R0's exact setup with LoRA off; W&B mobiuslab/miles-smoke/ni2fatd9): step 0 log-prob gap 0.1100 vs R0 0.1103, KL 0.0139 vs 0.0138, grad norm 1.008 vs 0.986; the first-sync bit-exact weight check passed and the NVMe optimizer step ran (rc 0).
• --check-only from this head (image built by Modal from 7b43958, recipe dsv41-flash-tb4-overfit16, 1 H100): check: ok. The CPU-only parse fails with "CUDA driver is a stub library", as it does on main (Megatron's parse_args needs CUDA).