freesolo-flash-dev 1.1.93__tar.gz → 1.2.7__tar.gz

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.1.93
+Version: 1.2.7
 Summary: Flash — managed LoRA post-training (SFT/GRPO/OPD) for Freesolo environments, driven by the `flash` CLI
 Project-URL: Homepage, https://github.com/freesolo-co/flash
 Project-URL: Repository, https://github.com/freesolo-co/flash

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/flash/runner/__init__.py

@@ -751,7 +751,6 @@ from flash.runner.artifacts import (  # noqa: E402,F401
     _assign_resolved_env_sha,
     _environment_artifact_repo_name,
     _file_digest,
-    _pin_env_sha_with_reason,
     artifact_namespace,
     flash_code_prefix,
     managed_hf_repo_for_environment,

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/flash/runner/artifacts.py

@@ -77,26 +77,13 @@ def _assign_managed_hf_repo(spec: JobSpec) -> JobSpec:
     return runner.JobSpec.from_dict(d)
 
 
-def _pin_env_sha_with_reason(spec: JobSpec) -> tuple[JobSpec, str]:
-    """Pin env ref->SHA, returning the spec and why the pin failed (or "" when it did not).
-
-    One resolve, two answers. The pin itself is best-effort and its caller only ever sees the
-    ABSENCE of a sha, which is the same observation for four different causes -- a ref that does
-    not exist, a rate limit, an outage, a private repo the token cannot read -- each needing a
-    different fix. GitHub already answered with which one; this carries that answer out instead of
-    discarding it.
-
-    Returning the reason rather than re-resolving on the failure path matters for correctness, not
-    just cost: a second call can succeed where the first failed (a rate-limit window that reset, a
-    blip that cleared), and a caller that has already committed to rejecting would then report an
-    empty reason for a ref that just resolved fine. The reason must describe the attempt whose
-    result is actually being used.
-    """
+def _assign_resolved_env_sha(spec: JobSpec) -> JobSpec:
+    """Pin env ref->SHA once so N workers don't fan-out N GitHub API calls (secondary rate-limit). Best-effort."""
     import logging
 
     env_id = spec.environment.id
     if not env_id or spec.environment.resolved_sha:
-        return spec, ""
+        return spec
     try:
         from flash.envs.loader import (
             _parse_github_environment_ref,
@@ -110,20 +97,15 @@ def _pin_env_sha_with_reason(spec: JobSpec) -> tuple[JobSpec, str]:
         )
         parsed = _parse_github_environment_ref(ref_str)
         if parsed is None:
-            return spec, ""
+            return spec
         sha = _resolve_ref_sha(parsed, timeout=10.0, max_rate_limit_retries=0)
     except Exception as e:
         logging.getLogger(runner.__name__).warning(
             "resolve-once: could not pin env ref->sha for %r (%s); worker will resolve", env_id, e
         )
-        return spec, str(e).strip()
+        return spec
     if not sha:
-        return spec, ""
+        return spec
     d = spec.to_internal_dict()
     d["environment"] = {**d["environment"], "resolved_sha": sha}
-    return runner.JobSpec.from_dict(d), ""
-
-
-def _assign_resolved_env_sha(spec: JobSpec) -> JobSpec:
-    """Pin env ref->SHA once so N workers don't fan-out N GitHub API calls (secondary rate-limit). Best-effort."""
-    return _pin_env_sha_with_reason(spec)[0]
+    return runner.JobSpec.from_dict(d)

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/flash/runner/preparation.py

@@ -619,20 +619,14 @@ def _profile_producer_version() -> str:
 
 
 def _require_pinned_profile_environment(spec: JobSpec) -> JobSpec:
-    pinned, reason = _runner()._pin_env_sha_with_reason(spec)
+    pinned = _runner()._assign_resolved_env_sha(spec)
     if not pinned.environment.id:
         raise _runner().WorkloadProfileUnavailable(
             "sft workload profiling requires an environment id"
         )
     if not pinned.environment.resolved_sha:
-        # the pin is best-effort, so reaching here says only that it did not happen -- and the four
-        # causes (a ref that does not exist, a rate limit, an outage, a private repo the token
-        # cannot read) need four different fixes. GitHub already answered with which one; name it
-        # instead of describing the missing pin it produced.
-        detail = f": {reason}" if reason else ""
         raise _runner().WorkloadProfileUnavailable(
-            f"sft workload profiling requires a pinned environment package revision, but "
-            f"{spec.environment.id!r} could not be resolved{detail}"
+            "sft workload profiling requires a pinned environment package revision"
         )
     return pinned
 

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.1.93"
+version = "1.2.7"
 description ="Flash — managed LoRA post-training (SFT/GRPO/OPD) for Freesolo environments, driven by the `flash` CLI"
 readme = "README.md"
 # RunPod Flash supports Python 3.11-3.12 (not 3.13 yet).
@@ -137,7 +137,7 @@ artifacts = ["flash/py.typed"]
 # auto-published by .github/workflows/publish-dev.yml: a push to `dev` whose version isn't yet on
 # PyPI publishes it; pushes that don't change it find it already published and no-op. So cutting a
 # dev-channel release = bump this (and [project].version) and merge to `dev`.
-version = "1.1.93"
+version = "1.2.7"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/tests/_helpers/profile.py

@@ -184,12 +184,7 @@ def satisfy_sft_profile(runner, monkeypatch, spec: Any):
     # Hub/model round-trips are replaced in place. Each has dedicated coverage; reaching the network
     # here would make every sft submit test an integration test.
     monkeypatch.setattr(runner, "_resolve_model_revision", lambda s, **_kw: _pin_model_revision(s))
-    # both entry points, because they are two views of one resolve and different callers take
-    # different ones: the sft profile gate needs the failure REASON and so calls
-    # `_pin_env_sha_with_reason`, while submit only wants the spec. stubbing one leaves the other
-    # reaching real GitHub for a `github:owner/repo` ref that does not exist.
     monkeypatch.setattr(runner, "_assign_resolved_env_sha", _pin_env_revision)
-    monkeypatch.setattr(runner, "_pin_env_sha_with_reason", lambda s: (_pin_env_revision(s), ""))
     stub_revision_geometry(monkeypatch)
 
     pinned = _pin_env_revision(_pin_model_revision(spec))

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/tests/test_provider_routing.py

@@ -641,104 +641,6 @@ def test_sft_submission_fails_closed_when_the_environment_cannot_be_pinned(orch,
     assert persisted == []
 
 
-@pytest.mark.parametrize(
-    ("github_error", "expected"),
-    [
-        pytest.param(
-            'GitHub environment request failed (422): {"message":"No commit found for SHA: main"}',
-            "No commit found for SHA: main",
-            id="wrong-ref",
-        ),
-        pytest.param(
-            "GitHub API rate limit exceeded (403)",
-            "rate limit",
-            id="rate-limit",
-        ),
-        pytest.param(
-            'GitHub environment request failed (404): {"message":"Not Found"}',
-            "404",
-            id="private-or-missing",
-        ),
-    ],
-)
-def test_unpinnable_sft_environment_reports_githubs_own_diagnosis(
-    orch, monkeypatch, github_error, expected
-):
-    """The pin failure must name WHY, not just that a revision is missing.
-
-    `@main` against a master-default repo made GitHub answer "No commit found for SHA: main" -- the
-    ref, the cause and the fix in one line. The plane had that string and threw it away, reporting
-    only "sft workload profiling requires a pinned environment package revision", which names none
-    of the three. A rate limit, an outage and a private repo the token cannot read all produce the
-    identical missing pin and all need different fixes, so the generic text sends every one of them
-    to the wrong remedy.
-    """
-    from dataclasses import replace
-
-    import flash.envs.loader as env_loader
-    from flash.providers import allocator
-
-    def fail(_parsed, *_args, **_kwargs):
-        raise RuntimeError(github_error)
-
-    monkeypatch.setattr(env_loader, "_resolve_ref_sha", fail)
-    monkeypatch.setattr(
-        orch,
-        "_resolve_model_revision",
-        lambda spec, **_kw: replace(spec, model_revision="b" * 40, model_revision_auto=True),
-    )
-    monkeypatch.setattr(
-        allocator, "allocate", lambda *a, **k: pytest.fail("allocated without a profile")
-    )
-
-    with pytest.raises(orch.WorkloadProfileUnavailable) as excinfo:
-        orch.submit_job(_public_spec())
-
-    message = str(excinfo.value)
-    assert expected in message, message
-    # the ref itself has to appear too: the operator has to know WHICH environment to fix.
-    assert "github:owner/repo@main:env/environment.py" in message
-
-
-def test_the_reported_reason_describes_the_resolve_that_was_actually_used(orch, monkeypatch):
-    """One resolve, so a retry that would have succeeded cannot blank out the reason.
-
-    Diagnosing by re-resolving is wrong twice over. A transient cause (a rate-limit window that
-    resets, a blip that clears) can succeed on the second call, and the caller -- already committed
-    to rejecting on the FIRST result -- would then report no reason at all, landing back on the
-    generic message this change exists to replace. It also doubles GitHub calls on every failing
-    submit, against the same rate limit that is one of the causes.
-    """
-    from dataclasses import replace
-
-    import flash.envs.loader as env_loader
-    from flash.providers import allocator
-
-    calls = []
-
-    def flaky(_parsed, *_args, **_kwargs):
-        calls.append(1)
-        if len(calls) == 1:
-            raise RuntimeError("GitHub API rate limit exceeded (403)")
-        return "b" * 40  # a second attempt SUCCEEDS; the rejection is already decided
-
-    monkeypatch.setattr(env_loader, "_resolve_ref_sha", flaky)
-    monkeypatch.setattr(
-        orch,
-        "_resolve_model_revision",
-        lambda spec, **_kw: replace(spec, model_revision="b" * 40, model_revision_auto=True),
-    )
-    monkeypatch.setattr(
-        allocator, "allocate", lambda *a, **k: pytest.fail("allocated without a profile")
-    )
-
-    with pytest.raises(orch.WorkloadProfileUnavailable) as excinfo:
-        orch.submit_job(_public_spec())
-
-    assert "rate limit" in str(excinfo.value), str(excinfo.value)
-    assert calls == [1], f"the pin must be resolved exactly once, got {len(calls)} calls"
-
-
 def test_lifecycle_fallback_pin_is_persisted_for_recovery(orch, monkeypatch):
     """A pin the lifecycle fallback recovers must survive a control-plane restart.
 

{freesolo_flash_dev-1.1.93 → freesolo_flash_dev-1.2.7}/uv.lock

@@ -1052,7 +1052,7 @@ wheels = [
 
 [[package]]
 name = "freesolo-flash"
-version = "1.1.93"
+version = "1.2.7"
 source = { editable = "." }
 
 [package.optional-dependencies]