[Lilo] Miles LoRA backend: DP>1 support + validated Qwen3.5-9B 16k on LongRLVR

[micahtyong]

Summary

Backend + definition changes needed for the Lilo-vs-Miles 9B LongRLVR parity runs (results in the W&B report Validating Multi-client LoRA on Lilo). This PR is the library surface those runs validated.

Miles LoRA backend: allow data parallelism (world_size % tp == 0 instead of == tp)

• MilesBackendConfig.data_parallel_size = world_size // tensor_model_parallel_size.
• Miles shards a forward_backward batch across DP ranks, so ragged batches are padded rather than rejected:

  pad_slot_rows(slot_rows, multiple=dp)  # appends zero-weight 2-token rows
  # pad row: {"tokens": t[:2], "target_len": 1, "target_tokens": t[1:2], "weights": [0.0], ...}

  MilesCommandBackend.forward_backward trims the padded outputs back to the real row count.
• Checkpoint metadata topology now records data_parallel_size; _validate_checkpoint treats a missing key in pre-existing checkpoints as 1, so old TP-only checkpoints still restore on an unchanged deployment while DP mismatches are rejected.

Definitions

• qwen3_5_9b_miles_lora_16k (8×H100, TP8, catalog-visible) and qwen3_5_9b_miles_lora_16k_dp2 (TP4×DP2, hidden) — identical except id/TP/visibility; the DP2 one is what the parity runs used (81 s/step vs 107 s on TP8).
• serve.py/app.py: APP_NAME = os.environ.get("LILO_APP_NAME", "lilo") so a second app (e.g. lilo-dp2) can be deployed alongside prod without code changes.

Tests: DP config acceptance/rejection, pad/trim behavior, topology metadata, legacy-checkpoint compatibility.

Link to Devin session: https://modal.devinenterprise.com/sessions/8d6804dd317c41329d77eb2c427dda6f
Open in Devin Desktop: https://modal.devinenterprise.com/desktop/session/8d6804dd317c41329d77eb2c427dda6f?variant=devin
Requested by: @micahtyong

---

[devin-ai-integration]

🤖 Devin AI Engineer

I'll be helping with this pull request! Here's what you should know:

✅ I will automatically:

• Address comments on this PR that start with 'DevinAI' or '@devin'.
• Look at CI failures and help fix them

Note: I can only respond to comments from users who have write access to this repository.

⚙️ Control Options:

• Disable automatic comment, CI, and merge conflict monitoring

[micahtyong]

/devin review

[devin-ai-integration]

Starting Devin Review.

[devin-ai-integration]

Devin Review found 8 potential issues.

[micahtyong]

When DP > 1 (two replicas of the model), Miles will split the batch into one shard per DP rank. Since every rank must receive at least one row, this method is used to ensure the number of rows is divisible by DP. Without this, a DP rank would hang

[micahtyong]

This config is validated with an actual e2e miles vs lilo run here

[micahtyong]

Going to merge this in @kailash109. Next one is #31 which is where we added context parallelism support for up to 128K.