Support independent input and output steps by alxmrs · Pull Request #860 · m2lines/Samudra

alxmrs · 2026-08-25T23:05:38Z

hist has been used to cover two concepts: the number of input steps and the number of output steps. These have always been decided together. This PR separates these two concepts so we can change them independently. Doing so, to me, clarifies what concept we actually need in various parts of how hist was used -- code paths either need input or output steps, and this PR makes use of the proper concept by name now instead of implicitly. I find that this makes reading various data flow paths more understandable than before.

🤖 below

Summary

add backwards-compatible data.output_steps configuration, defaulting to hist + 1
support rolling input histories when a model emits fewer timesteps than it consumes
anchor residual predictions to the most recent matching input span
separate input/output timestep handling in validation, rollout metrics, and Zarr writing
cover the 2-input / 1-output dataset, residual rollout, Trainer, validation, and writer paths

This enables hist: 1, output_steps: 1: two input states predict one future state, and each autoregressive call shifts [x(t-1), x(t)] to [x(t), x_hat(t+1)].

Existing configs omit output_steps and retain their current hist + 1 output blocks.

Validation

uvx pre-commit run --all-files
.venv/bin/pytest -m 'not manual and not cuda' -q
- 417 passed, 2 skipped, 67 deselected, 10 xfailed

alxmrs

Initial review to agents.

alxmrs · 2026-08-28T18:25:42Z

        self._tensor_map = tensor_map
        self.num_prognostic_channels = num_prognostic_channels
-        self.hist = hist
+        self.input_hist = hist


Maybe we can make this input_steps that is hist+1 to be consistent with the accounting of output_steps, rather than use output_hist which is a weird concept.

alxmrs · 2026-08-28T23:47:58Z

+        ),
+    )
+    input_steps: int | None = Field(
+        default=None,


This should be defaulted to 2 IMO, since his was defaulted to 1 before.

alxmrs · 2026-08-31T17:12:06Z

@fomo-bot may I have your review? If you think the PR is ready for merge, will you approve it?

fomo-bot

Reviewed current head 4ccafb7. The independent input/output step refactor looks ready to merge after focused local validation and green current-head CI.

fomo-bot · 2026-08-31T17:40:28Z

Approved PR #860.

I reviewed current head 4ccafb76aef3b613198e9c804f721597a58812bb, including live PR metadata, full diff, review threads, CI, and local validation. I did not find blocking issues. The remaining unresolved review threads are outdated, and their requested direction is reflected in the current code.

Validation:

git diff --check origin/main...HEAD passed.
Focused local suite passed: 249 passed, 65 deselected, 3 warnings.
The failed GPU workflow was an EC2 InsufficientInstanceCapacity launch failure; I reran failed jobs and the GPU run passed.
Final GitHub state: reviewDecision=APPROVED, mergeable=MERGEABLE, mergeStateStatus=CLEAN.

Debug info

Repository: m2lines/Samudra
Trigger: Support independent input and output steps #860 (comment)
Request ID: 1a59f9a2-a81d-4dda-a0fd-2f6eee7d2f88
Accepted: true

alxmrs · 2026-09-01T22:41:10Z

Hey @codex, will you review this PR?

chatgpt-codex-connector · 2026-09-01T22:44:05Z

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: 4ccafb76ae

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

alxmrs · 2026-09-02T17:59:14Z

@codex may I have your review on this PR?

chatgpt-codex-connector · 2026-09-02T18:07:24Z

Codex Review: Didn't find any major issues. More of your lovely PRs please.

Reviewed commit: cfbf15cfeb

ℹ️ About Codex in GitHub

Codex has been enabled to automatically review pull requests in this repo. Reviews are triggered when you

Open a pull request for review
Mark a draft as ready
Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".

amogh-gulati

Looks good! just one small comment which came up

amogh-gulati · 2026-09-04T21:09:37Z

🐑 : This is still old explanation right? where the inputs and outputs are coupled

Good catch—the docstring was still describing only the coupled legacy behavior. I updated it to define windows in terms of independent input_steps and output_steps, including explicit 2-in/2-out and 2-in/1-out examples (2f49ac9).

github-project-automation Bot added this to Samudra Planning Aug 25, 2026

github-project-automation Bot moved this to Backlog in Samudra Planning Aug 25, 2026

alxmrs commented Aug 28, 2026

View reviewed changes

Comment thread src/samudra/models/base.py Outdated

alxmrs commented Aug 28, 2026

View reviewed changes

alxmrs changed the title ~~Support independent input and output horizons~~ Support independent input and output steps Aug 29, 2026

alxmrs commented Aug 29, 2026

View reviewed changes

Comment thread src/samudra/config.py Outdated

alxmrs requested a review from amogh-gulati August 31, 2026 17:11

fomo-bot approved these changes Aug 31, 2026

View reviewed changes

alxmrs force-pushed the feature/independent-prediction-horizon branch from 4ccafb7 to cfbf15c Compare September 1, 2026 22:41

alxmrs force-pushed the feature/independent-prediction-horizon branch from cfbf15c to 2a54ac2 Compare September 2, 2026 18:18

alxmrs added 2 commits September 4, 2026 10:44

Support independent input and output horizons

d60895c

Fix canonical dataset horizon test

a66e9c6

alxmrs force-pushed the feature/independent-prediction-horizon branch from 662a87f to a66e9c6 Compare September 4, 2026 17:48

amogh-gulati approved these changes Sep 4, 2026

View reviewed changes

alxmrs added 2 commits September 4, 2026 15:51

Document independent inference horizons

2f49ac9

Merge branch 'main' into feature/independent-prediction-horizon

455fd66

alxmrs enabled auto-merge September 4, 2026 23:04

alxmrs added this pull request to the merge queue Sep 4, 2026

Merged via the queue into main with commit 55fe01d Sep 4, 2026
20 checks passed

alxmrs deleted the feature/independent-prediction-horizon branch September 4, 2026 23:35

github-project-automation Bot moved this from Backlog to Done in Samudra Planning Sep 4, 2026

Uh oh!

Conversation

alxmrs commented Aug 25, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Summary

Validation

Uh oh!

alxmrs left a comment

Choose a reason for hiding this comment

Uh oh!

alxmrs Aug 28, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

alxmrs Aug 28, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

alxmrs commented Aug 31, 2026

Uh oh!

fomo-bot left a comment

Choose a reason for hiding this comment

Uh oh!

fomo-bot commented Aug 31, 2026

Uh oh!

alxmrs commented Sep 1, 2026

Uh oh!

chatgpt-codex-connector Bot commented Sep 1, 2026

Uh oh!

alxmrs commented Sep 2, 2026

Uh oh!

chatgpt-codex-connector Bot commented Sep 2, 2026

Uh oh!

amogh-gulati left a comment

Choose a reason for hiding this comment

Uh oh!

amogh-gulati Sep 4, 2026

Choose a reason for hiding this comment

Uh oh!

alxmrs Sep 4, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

alxmrs commented Aug 25, 2026 •

edited

Loading