Fix imports to be global not relative for patches#76
Open
joyliu-q wants to merge 29 commits into
Open
Conversation
Adds memory, cloud, and region fields to ModalConfig and wires them into the @app.function decorator in slime/modal_train.py so configs can request per-container memory limits and pin cloud/region. The 355B base config (glm47_355b_a32b) sets memory=(1024, 2 TiB) so the trainer host has headroom for the FP32 delta baseline copy without OOM. The non-colocated variant now inherits via `modal = _base.modal` instead of redeclaring, and its docstring is refreshed (EP=16 is unchanged from base, not reduced). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Plan generated via gen-plan with 3 Codex convergence rounds. Covers miles/ Modal launcher mirroring slime/ contract, Qwen3-4B LoRA smoke test, and 7 acceptance criteria. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Implements the miles/ launcher directory mirroring the slime/ contract: - configs/base.py: ModalConfig + MilesConfig with reflection-based CLI, JSON config field support, and NCCL_NVLS_ENABLE=1 in base env - modal_train.py: 5 Modal functions (list_configs, download_model, prepare_dataset, convert_checkpoint, train) - modal_helpers/utils.py: cluster context, command building, config prep - configs/qwen3_4b_lora_smoke.py: Qwen3-4B LoRA verification config (num_rollout=2, bridge mode, 4x H200 colocated) Docker image: radixark/miles:dev-202604101227 Volumes: huggingface-cache (shared), miles-data, miles-checkpoints Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Fix prepare_data() to use zhuzilin/dapo-math-17k and zhuzilin/aime-2024 (matching slime sibling configs), add os.makedirs before download - Expand README with prerequisites, secrets, step-by-step workflow, config authoring guide, YAML/JSON fields, and patch injection docs (matching slime/README.md quality) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The train function now only attaches wandb-secret when the experiment config has use_wandb=True. The smoke-test config (which disables WandB) no longer depends on this secret being present. Also documents required Modal secrets in README prerequisites. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Modal evaluates decorator args locally and remotely, so conditional secrets cause dependency mismatch errors. Always attach wandb-secret unconditionally, matching slime's pattern. Also picks up user config changes (H100, 8 GPUs per node). Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- configs/kimi_k25_fullparam_smoke.py: 8x H200 colocated, bridge mode, freezes everything except layer 60 for smoke testing. Depends on the K2.5 weight-update fixes landed in the miles repo. - configs/kimi_k25_lora.py: LoRA training on top of smoke config. Targets MLA attention projections and MLP linear layers (rank 32, alpha 32). - test_bridge.py: single-GPU Modal script that exercises bridge registration, provider construction, and model instantiation. Used for fast iteration when debugging bridge/provider issues without a full multi-node run. - configs/base.py: add memory and image_env fields to ModalConfig so configs can request per-container RAM (K2.5 needs ~1.8 TiB) and bake LD_LIBRARY_PATH into the image. - modal_train.py: bump base image to radixark/miles:dev (rolling), always prepend system libs to LD_LIBRARY_PATH, and add a convert_kimi_int4_to_bf16 helper to prepare the bf16 checkpoint from moonshotai/Kimi-K2.5. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Drop sglang_pynccl_nonfatal.patch.py from image_run_commands (patch was removed from miles repo). - Remove the commented-out INT4 QAT __init__ override and the unused _MILES_PR896_SHA constant. - Drop the no-op prepare_data config-patching branch — the checkpoint on the volume is already in the right shape. - Tighten image_run_commands comments to name the actual issue each step addresses instead of paraphrasing. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- Phase 1 target_modules: drop linear_q_proj (unused — K2.5 has q_lora_rank so MLA decomposes into down+up; linear_q_proj only exists when q_lora_rank is None). Drop linear_fc1/linear_fc2 — the sparse-expert + compressed- tensors + LoRA path is the least-tested. Keep MLA decomposition (q_down/q_up/kv_down/kv_up) + o_proj. Gives 5 × 60 = 300 adapters instead of ~138K when targeting all expert MLPs. - Enable OPEN_TRAINING_INT4_FAKE_QAT_FLAG=1 + OPEN_TRAINING_INT4_GROUP_SIZE=32 so training forward sees INT4-quantized base weights, matching SGLang's INT4 inference. Without this the LoRA adapters would learn to compensate for a BF16↔INT4 numerical gap that doesn't exist at inference time. LoRA doesn't optimize the base weights so the dequant scratch buffer fits in H200 memory (unlike the full-param smoke where it OOMed). - Drop the explicit save path; save/save_interval are unset in the parent smoke config, so no checkpoint is written. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Switch target_modules back to the full general-case set: linear_q_down_proj, linear_q_up_proj, linear_kv_down_proj, linear_kv_up_proj, linear_proj, linear_fc1, linear_fc2 Per https://thinkingmachines.ai/blog/lora/ attention-only LoRA significantly underperforms MLP-only, and MLP+attention ≈ MLP-only. If the full path works end-to-end, narrower scopes follow trivially — so exercise the hardest code path first: MLA attention LoRA + dense MLP LoRA + shared expert LoRA + sparse expert (GroupedMLP) LoRA on compressed-tensors experts. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Contributor
|
I think we should make all these paths relative with sth like |
Introduces miles/configs/model_configuration.py with a ModelConfiguration base class (model_name, model_path, download_model) and a KimiK25 subclass whose download_model() shells out via `modal run` to the existing download_model and convert_kimi_int4_to_bf16 Modal app functions. Adds miles/test_model_configuration.py covering import, construction, the no-CLI-arg-leak guard (via _MILES_SKIP), and NotImplementedError on the base. Round 0 of the surge-kimi-plan RLCR loop (live residuals only; the 1-step smoke goal was met by an external 8x H200 MLA 8k x 256 run on 2026-04-21). Companion updates to _MILES_SKIP and the two experiment scanners are already in 🆕 v0. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
joyliu-q
force-pushed
the
joy/improve-nan-miles-refactor
branch
from
8cd4a8a to
2914c58
Compare
joyliu-q
commented
Apr 21, 2026
| _MODAL_TRAIN_PATH = _MILES_ROOT / "modal_train.py" | ||
|
|
||
|
|
||
| class KimiK25ModelConfiguration(ModelConfiguration): |
Contributor
Author
There was a problem hiding this comment.
TODO: Add more model configs and port other model configs over to use this pattern.
nanjiangwill
force-pushed
the
nan/miles-refactor
branch
2 times, most recently
from
4e5df13 to
f19d3af
Compare
![devin-ai-integration[bot]](https://avatars.githubusercontent.com/in/811515?s=60&v=4){kind=small}
| # docs/agent-modal-training.md. | ||
| env = {**os.environ, "EXPERIMENT_CONFIG": "kimi_k25"} | ||
| subprocess.run( | ||
| ["modal", "run", f"{_MODAL_TRAIN_PATH}::download_model"], |
Contributor
There was a problem hiding this comment.
🔴 KimiK25ModelConfiguration.download_model references non-existent modal function
miles/configs/model_configuration.py:49 invokes modal run modal_train.py::download_model, but the download_model function was removed in this PR and replaced with prepare_model (miles/modal_train.py:127). When KimiK25ModelConfiguration.download_model() is called (e.g. from the test file or agent workflows), the subprocess will fail with a Modal "function not found" error because no function named download_model exists in the app.
Suggested change
| ["modal", "run", f"{_MODAL_TRAIN_PATH}::download_model"], | |
| ["modal", "run", f"{_MODAL_TRAIN_PATH}::prepare_model"], | |
Was this helpful? React with 👍 or 👎 to provide feedback.
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
3 participants
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.
Will add other fixes if I needed to make any to make this run out of the box!