[zhou9402]

Ports miles' FLOPs model / MFU metric and the metric-history CI gate (SQLite store + declarative register_ci_gate).

Source: radixark/miles@ba28991834. Port is additive (new modules + new add_*_arguments() groups); slime-origin files are not restructured so future [Sync] Mirror Slime PRs stay clean. One commit per sub-PR; review each commit separately.

Commits:

• 89ddd2a F10-0 (V5): FLOPs oracle verifier
• 2cea872 F10-1 (A1a): full HF-config FLOPs model in vime
• add5737 F10-2 (A1b): MFU metric + --mfu-peak-tflops + device table
• fa0ef42 F10-3: metric-history SQLite storage layer
• 8b75e2c F10-4: metric-history gate core (selection/constraints/evaluate_gate)
• fc8fd9a F10-5: register_ci_gate declaration parsing
• 87c5a98 F10-6 (V1): metric capture backend + gate runner + e2e verifier
• 06e4cb4 F10-7: Buildkite wiring for the FLOPs oracle and metric-history gate

Verification: tests/ci/flops_oracle/ (V5) checks the FLOPs model against a hand-computed oracle; tests/ci/metric_history/ (V1) runs the gate end to end on a synthetic history. GPU suite baseline persistence still needs one real Buildkite run.

Merge order of the miles-port series: F10 → F02 → F06 → F11a → F11b → F05 → F09 → F04 → F03 → F07 → F08 → F01. train.py / vime/utils/arguments.py are touched by several branches; later ones are rebased after each merge.

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request ports the MFU metric and metric-history regression gate tooling from the miles repository to vime, including the analytical FLOPs model, SQLite-backed metric-history store, and Buildkite CI integration. Feedback on these changes highlights three key improvements: moving the transformers import inside the exception handler in train_metric_utils.py to prevent training crashes, coercing step counters to standard Python integers in ci_metric_history.py to support numpy and PyTorch types, and automatically creating the parent directory for the SQLite database in sqlite_store.py to avoid operational errors.

[gemini-code-assist]

The from transformers import AutoConfig import is placed outside the try...except block. If transformers is not installed in the environment, this will raise an ImportError and crash the training process, violating the goal of the exception handler (accounting must never break training). Additionally, if hf_checkpoint is empty or not set, calling AutoConfig.from_pretrained will fail and log a noisy warning. Checking for an empty hf_checkpoint and moving the import inside the try block resolves both issues.

def _mfu_model_flops(hf_checkpoint: str):
    \"\"\"HF-config FLOPs model for MFU accounting, cached per checkpoint; None when unsizable.\"\"\"
    if not hf_checkpoint:
        return None

    try:
        from transformers import AutoConfig
        hf_config = AutoConfig.from_pretrained(hf_checkpoint, trust_remote_code=True)

[gemini-code-assist]

The check isinstance(step, int) will evaluate to False if step is a numpy integer (e.g., numpy.int64) or a PyTorch 0-dimensional tensor, which are very common types for step counters in training loops. This causes the step to be recorded as None (and serialized as null in JSONL), which subsequently triggers a SelectionError in the CI gate when evaluating steps: \"all\". Coercing the step to a standard Python integer using int() with a fallback to None ensures robust step capture.

Suggested change

	step = metrics.get(step_key)
	recorder.log(metrics, step if isinstance(step, int) and not isinstance(step, bool) else None)
	step = metrics.get(step_key)
	if step is not None and not isinstance(step, bool):
	try:
	step = int(step)
	except (TypeError, ValueError, RuntimeError):
	step = None
	else:
	step = None
	recorder.log(metrics, step)

[gemini-code-assist]

If the parent directory of the SQLite database file does not exist, sqlite3.connect will raise a sqlite3.OperationalError: unable to open database file and crash the CI gate. Automatically creating the parent directory if it does not exist makes the store initialization much more robust when custom paths are configured.

    def __init__(self, database: str = \":memory:\"):
        # check_same_thread is left at its default; the store is meant for the
        # single-threaded CI gate process and the test suite.
        if database != \":memory:\":
            import os
            parent = os.path.dirname(database)
            if parent:
                os.makedirs(parent, exist_ok=True)
        self._conn = sqlite3.connect(database)

[read-the-docs-community]

Documentation build overview

📚 vime | 🛠️ Build #35444823 | 📁 Comparing 06e4cb4 against latest (d7e40fe)

  🔍 Preview build  

59 files changed · + 1 added · ± 58 modified

+ Added

• developer_guide/mfu_metric_gate.html

± Modified

• genindex.html
• library.html
• search.html
• advanced/arch-support-beyond-megatron.html
• advanced/delta-weight-sync.html
• advanced/external-rollout-engines.html
• advanced/fault-tolerance.html
• advanced/low-precision.html
• advanced/megatron-config.html
• advanced/observability.html
• and 48 more...