Bump DeepSeek-V4.1-Flash to miles dev-202610082133 image

[zhouhelena1]

Closing this bc current image already has DeepSeek-V4.1 DSA attention kernel, so we already see the benefits of radixark/miles#3608. This new image does fix a TileLang forward bug at 16 heads with top-k 128, but this doesn't match our recipe, so I'd prefer sticking to the default image than switching to dev image.

---

Moves DeepSeek_V4_1_Flash_Recipe (8×8 H200) from the dedicated dsv41-h200-ea751aac8 image to the standard multi-arch dev image radixark/miles:dev-202610082133, pinned by digest. That image is miles f5757caab, which contains the radixark/miles#3605–#3634 kernel stack including #3608, with sglang-miles b84bec576 and Megatron-LM fd15ee20a. Other Miles recipes are unchanged.

Changes:

• Recipe: new image pin. Adds SGLANG_OPT_DEEPGEMM_HC_PRENORM=0 to match upstream's run_deepseek_v4_1.py, which turns off the compensated-MHC derived-weight cache that goes stale under online weight updates.
• DeepSeek patches: all 6 apply to the new image unchanged. Only their image/sglang revision docstrings and the renamed plugin path (deepseek_v41 → deepseek_v4_1) were updated.
• patch_zero_std_metrics: now also matches the new uniform_rewards layout from Fix uniform reward percentages for numeric rewards radixark/miles#3722. That upstream change still puts an int 1 and a float 1.0 in different buckets. Unit tests added.
• patch_advantage_distribution: finds log_utils.py at its new training_utils/metrics/ location (moved in training_utils: the shared layer for every training backend radixark/miles#3248). Before this fix it printed "not found" and was skipped.

Validation

3-step run of validate_model_configs.py check -m DeepSeek-V4.1-Flash -n 3 in helena-dev, compared with a run on the old image:

	old image messy-plate-d27a30d8cf5a	new image sophisticated-hoop-e728ce9825ea
reward (steps 1/2/3)	0.375 / 0.188 / 0.453	0.422 / 0.180 / 0.461
train↔rollout logprob abs diff	0.132 / 0.117 / 0.132	0.130 / 0.115 / 0.133
train↔rollout KL	0.056 / 0.043 / 0.057	0.056 / 0.042 / 0.058
grad norm	0.106 / 0.077 / 0.077	0.090 / 0.090 / 0.111
rollout time (s)	613 / 589 / 581	592 / 594 / 622
log_probs (s)	248 / 47 / 36	291 / 79 / 51
actor_train (s)	880 / 373 / 395	1018 / 341 / 356
checkpoint save (s)	412	436

Rollout traces for all 3 steps (128 samples each) look like the old image's. Responses are coherent and end with \boxed{} answers, and the number truncated at 4096 tokens is about the same (66/105/69 vs 71/103/70). The new image reports ppo_kl and pg_clipfrac as exactly 0. With one optimizer step per rollout that means the training forward pass now reproduces the log_probs pass bit for bit, where the old image differed by about 1e-4.

Wall time is roughly unchanged for this setup. In steps 2–3, actor_train is about 9% faster, which fits #3608's faster DSA attention. But optimizer step (about 4.5 min) dominates training time and the stack doesn't change it, while log_probs is somewhat slower. Pre-existing gaps that are the same on both images: dashboard stage labels for rollout substeps, and train_async substep timing.

Locally: ruff, format, compileall, pre-commit, and the tests/ files pass. Two dashboard test files, test_dashboard_ingestion_auth.py and test_dashboard_metrics.py, hang locally; this PR doesn't touch them.

Checklist

• Example is documented with comments throughout, in a Literate Programming style.
• Example does not require third-party dependencies to be installed locally
• Example pins its dependencies
  • Example pins container images to a stable tag, not a dynamic tag like latest
  • Example specifies a python_version for the base image, if it is used
  • Example pins all dependencies to at least minor version, ~=x.y.z or ==x.y
  • Example dependencies with version < 1 are pinned to patch version, ==0.y.z

[github-actions]

Documentation preview: https://modal-labs-training-gym--training-gym-previews-preview-r-78da35.modal.run/689/docs