freesolo-flash-dev 1.0.75__tar.gz → 1.0.76__tar.gz

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: freesolo-flash-dev
-Version: 1.0.75
+Version: 1.0.76
 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-1.0.75 → freesolo_flash_dev-1.0.76}/flash/cli/TRAINING.md

@@ -848,7 +848,7 @@ with no reward to design. It supports `epochs` like SFT/GRPO and produces a LoRA
 
 - **Pick the teacher with `[train] teacher_model`; the key stays managed.** The teacher defaults to
   the managed **GLM 5.2** and is selectable from a fixed, managed allow-list:
-  `glm-5.2` (default), `deepseek-v4-pro`, `kimi-k2.6`. Every option is
+  `glm-5.2` (default) or `kimi-k2.6`. Every option is
   a Fireworks-hosted model reached with the platform's own key, so there is nothing to export or
   declare — an opd run submits like any other, and a `FIREWORKS_API_KEY` in your shell is ignored.
   Arbitrary bring-your-own teacher models or keys are not supported (the allow-list is curated to
@@ -883,7 +883,7 @@ epochs = 1
 max_examples = 2
 lora_rank = 32
 # teacher_model = "glm-5.2"                             # managed teacher to distil from; one of
-#                                                       # glm-5.2 (default) | deepseek-v4-pro | kimi-k2.6
+#                                                       # glm-5.2 (default) | kimi-k2.6
 #                                                       # (key stays managed)
 # kl_penalty_coef = 1.0                                 # reverse-KL scale
 ```

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/flash/cli/env_setup.py

@@ -667,7 +667,7 @@ def cmd_env_setup(args) -> int:
             f"{max_examples_line}"
             "lora_rank = 32\n"
             '# teacher_model = "glm-5.2"   # teacher to distil from: glm-5.2 (default) |\n'
-            "#                             # deepseek-v4-pro | kimi-k2.6 (key stays managed)\n"
+            "#                             # kimi-k2.6 (key stays managed)\n"
             "# GPU and HF artifacts are managed automatically by the platform: the GPU is\n"
             "# the cheapest fitting managed class, and artifacts live in a private environment-scoped repo.\n"
         )

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/flash/engine/recipe.py

@@ -26,8 +26,8 @@ class TeacherModel:
     display_name: str
     # Fireworks serverless list price, $/1M tokens as (input, output). OPD echo-scores completions
     # (max_tokens=0) so only the INPUT column is billed / feeds the estimate, but both are kept so a
-    # mispriced row is obvious. NOTE: the non-GLM entries are best-effort estimates for these next-gen
-    # serverless models — confirm against the live fireworks.ai pricing before relying on the
+    # mispriced row is obvious. note: the non-glm entry is a best-effort estimate for this next-gen
+    # serverless model — confirm against the live fireworks.ai pricing before relying on the
     # `flash train` cost quote.
     usd_per_1m: tuple[float, float]
     supports_images: bool = False
@@ -49,12 +49,6 @@ TEACHER_MODELS: dict[str, TeacherModel] = {
         display_name="GLM 5.2",
         usd_per_1m=(1.40, 4.40),
     ),
-    "deepseek-v4-pro": TeacherModel(
-        alias="deepseek-v4-pro",
-        model_id="accounts/fireworks/models/deepseek-v4-pro",
-        display_name="DeepSeek V4 Pro",
-        usd_per_1m=(1.74, 3.48),
-    ),
     "kimi-k2.6": TeacherModel(
         alias="kimi-k2.6",
         model_id="accounts/fireworks/models/kimi-k2p6",

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/flash/engine/worker/heartbeat.py

@@ -37,11 +37,14 @@ _HB_SETUP_LIVENESS_STAGES = frozenset(
         "sft_configuring",
         "opd_configuring",
         "opd_filtering_prompts",
+        "opd_prompt_scan",
+        "opd_image_prep",
         "sft_initializing",
         "rl_initializing",
         "opd_initializing",
         "sft_finalizing",
         "rl_finalizing",
+        "opd_finalizing",
     }
 )
 # Mid-training the per-step checkpoint upload runs SYNCHRONOUSLY on the trainer thread (dev #445),
@@ -57,11 +60,11 @@ _HB_TIGHT_LIVENESS_STAGES = _HB_SETUP_LIVENESS_STAGES | _HB_UPLOAD_LIVENESS_STAG
 # latest per-step GRPO backlog, exposed so a top-level error heartbeat can preserve it
 # for `flash runs log -f` when a short run raises before the throttled rl_step ping committed
 LATEST_GRPO_METRICS_LAST: list = []
-# Throttled to avoid blowing the 128/hr HF commit cap; terminal transitions are never throttled. Every
-# tight-liveness stage is throttled (⊂) PLUS the per-step training stages: opd_filtering_prompts alone
-# emits a REAL (non-liveness) heartbeat every scan tick — ~120/hr on a large split before model load —
-# so unthrottled the setup stages blow the cap; throttle them exactly like their sft_pretokenizing
-# analogue (codex[bot]). checkpoint_uploading keepalive re-emits every 30s too, so it MUST be throttled.
+# throttle these stages to protect the hf repository commit budget; terminal transitions are never throttled.
+# opd_filtering_prompts emits a real heartbeat on each scan tick, while opd_prompt_scan and
+# opd_image_prep emit one when progress advances. opd_finalizing emits one on every keepalive tick.
+# without throttling, these opd stages can make roughly 120 commit attempts per hour. tight-liveness,
+# per-step training, and checkpoint_uploading keepalive stages share the same throttle.
 _HB_THROTTLED_STAGES = _HB_TIGHT_LIVENESS_STAGES | frozenset({"rl_step", "sft_step", "opd_step"})
 _HB_TERMINAL_STAGES = frozenset({"done", "already_done"})
 # 600s -> ~6 commits/hr; keeps stall detector alive without hitting the HF commit cap.

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.0.75"
+version = "1.0.76"
 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).
@@ -123,7 +123,7 @@ 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 = "1.0.75"
+version = "1.0.76"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/tests/test_multimodal_training.py

@@ -618,7 +618,7 @@ def test_image_opd_preflight_validates_packaged_dataset_before_allocation(tmp_pa
         mm.preflight_validate_image_opd(unsupported)
 
 
-@pytest.mark.parametrize("teacher_model", ["", "glm-5.2", "deepseek-v4-pro"])
+@pytest.mark.parametrize("teacher_model", ["", "glm-5.2", "accounts/fireworks/models/glm-5p2"])
 def test_image_opd_preflight_requires_kimi_vision_teacher(tmp_path, teacher_model):
     root, _image = _package(tmp_path)
     env_file = root / "environment.py"

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/tests/test_opd.py

@@ -467,11 +467,8 @@ def test_opd_selects_managed_teacher_and_rejects_unknown():
             run_id="x",
         )
 
-    # Supported aliases parse and are stored as the canonical Fireworks model id.
+    # supported aliases parse and are stored as the canonical fireworks model id.
     assert _spec("kimi-k2.6").train.teacher_model == "accounts/fireworks/models/kimi-k2p6"
-    assert (
-        _spec("deepseek-v4-pro").train.teacher_model == "accounts/fireworks/models/deepseek-v4-pro"
-    )
     # A spaced / mixed-case form normalizes to the same model id.
     assert _spec("GLM 5.2").train.teacher_model == "accounts/fireworks/models/glm-5p2"
     # The raw Fireworks model id is also accepted (identity), including with stray surrounding
@@ -487,11 +484,15 @@ def test_opd_selects_managed_teacher_and_rejects_unknown():
     # Omitting/blank leaves it unset ("" => the worker uses the default GLM 5.2 teacher).
     assert _spec("").train.teacher_model == ""
 
-    # An unsupported teacher is rejected at parse time with a teacher-specific ConfigError.
+    # an unsupported teacher is rejected at parse time with a teacher-specific configerror.
     with pytest.raises(ConfigError, match="teacher_model"):
         _spec("gpt-5.5")
+    with pytest.raises(ConfigError, match="teacher_model"):
+        _spec("deepseek-v4-pro")
+    with pytest.raises(ConfigError, match="teacher_model"):
+        _spec("accounts/fireworks/models/deepseek-v4-pro")
     # qwen-3.7-max (on-demand only) and minimax-m3 (serverless chat, but its /completions echo
-    # endpoint OPD needs does not respond) are NOT allow-listed teachers, so both are rejected.
+    # endpoint opd needs does not respond) are not allow-listed teachers, so both are rejected.
     with pytest.raises(ConfigError, match="teacher_model"):
         _spec("qwen-3.7-max")
     with pytest.raises(ConfigError, match="teacher_model"):
@@ -652,17 +653,17 @@ def test_opd_filtering_stage_is_setup_not_training():
     )  # progress count doesn't flip it
 
 
-def test_opd_filtering_prompts_is_throttled_like_sft_pretokenizing():
-    """Regression (codex[bot], heartbeat.py): opd_filtering_prompts emits a REAL progress heartbeat
-    per liveness tick while it renders+tokenizes the whole split. Unthrottled that is one HF commit
-    per tick -- ~120/hr on a large split before model load, blowing the 128/hr commit cap. It must be
-    registered in BOTH heartbeat sets its SFT analogue sft_pretokenizing lives in: the throttle set
-    (bounds commit rate) and the setup-liveness set (keeps the tighter cold-start upload cadence)."""
+def test_opd_liveness_stages_are_throttled_at_setup_cadence():
+    """opd liveness threads must use the throttled setup-liveness cadence."""
     from flash.engine.worker.heartbeat import _HB_SETUP_LIVENESS_STAGES, _HB_THROTTLED_STAGES
 
+    opd_liveness_stages = {"opd_prompt_scan", "opd_image_prep", "opd_finalizing"}
+    assert opd_liveness_stages <= _HB_SETUP_LIVENESS_STAGES
+    assert opd_liveness_stages <= _HB_THROTTLED_STAGES
+
     assert "opd_filtering_prompts" in _HB_THROTTLED_STAGES
     assert "opd_filtering_prompts" in _HB_SETUP_LIVENESS_STAGES
-    # parity with the SFT pre-tokenize stage this mirrors (same dual membership).
+    # parity with the sft pre-tokenize stage this mirrors (same dual membership).
     assert "sft_pretokenizing" in _HB_THROTTLED_STAGES
     assert "sft_pretokenizing" in _HB_SETUP_LIVENESS_STAGES
 
@@ -1268,8 +1269,8 @@ def test_opd_teacher_rate_matches_fireworks_glm5p2_input_price():
 
 def test_opd_teacher_price_table_covers_every_allowlisted_teacher():
     """Every allow-listed teacher is priced by its exact row (pricing routes through resolve_teacher
-    over recipe.TEACHER_MODELS, so there is no unpriced teacher), the new teachers carry their own
-    input prices (not silently GLM-priced), and an unknown teacher falls back to the default rate."""
+    over recipe.TEACHER_MODELS, so there is no unpriced teacher), Kimi carries its own input price
+    (not silently GLM-priced), and an unknown teacher falls back to the default rate."""
     from flash.cost.facts import teacher_price_per_1m
     from flash.engine.recipe import TEACHER_MODELS
 
@@ -1277,11 +1278,10 @@ def test_opd_teacher_price_table_covers_every_allowlisted_teacher():
     for info in TEACHER_MODELS.values():
         assert teacher_price_per_1m(info.model_id) == info.usd_per_1m
 
-    # The two added teachers carry their own input prices (distinct from GLM's $1.40/M).
-    assert teacher_price_per_1m("accounts/fireworks/models/deepseek-v4-pro")[0] == 1.74
+    # kimi carries its own input price (distinct from glm's $1.40/m).
     assert teacher_price_per_1m("accounts/fireworks/models/kimi-k2p6")[0] == 0.95
-    # Removed teachers (qwen-3.7-max on-demand only; minimax-m3 no echo support) are unknown ids
-    # now -> priced defensively at the default (GLM) rate.
+    # removed teachers are unknown ids and price defensively at the default glm rate.
+    assert teacher_price_per_1m("accounts/fireworks/models/deepseek-v4-pro")[0] == 1.40
     assert teacher_price_per_1m("accounts/fireworks/models/qwen3p7-max")[0] == 1.40
     assert teacher_price_per_1m("accounts/fireworks/models/minimax-m3")[0] == 1.40
 
@@ -2212,17 +2212,21 @@ def test_resolve_opd_knobs_resolves_teacher_from_train(monkeypatch):
         )
         return opd_mod._resolve_opd_knobs()
 
-    # A friendly alias resolves to the provider model id.
+    # a friendly alias resolves to the provider model id.
     assert _knobs("kimi-k2.6").teacher_model == "accounts/fireworks/models/kimi-k2p6"
-    assert _knobs("deepseek-v4-pro").teacher_model == "accounts/fireworks/models/deepseek-v4-pro"
-    # Unset / blank / None -> the default GLM 5.2 teacher (historical behavior preserved).
+    # unset / blank / none -> the default glm 5.2 teacher (historical behavior preserved).
     assert _knobs("").teacher_model == "accounts/fireworks/models/glm-5p2"
     assert _knobs(None).teacher_model == "accounts/fireworks/models/glm-5p2"
-    # base_url is shared across every allow-listed teacher (one Fireworks endpoint + one managed key).
-    assert _knobs("deepseek-v4-pro").teacher_base_url == opd_mod.RECIPE.opd.teacher_base_url
-    # An unsupported teacher fails loudly on the worker (defensive guard, mirrors the kl_coef check).
-    with pytest.raises(RuntimeError, match="teacher_model"):
-        _knobs("gpt-5.5")
+    # base_url is shared across every allow-listed teacher (one fireworks endpoint + one managed key).
+    assert _knobs("kimi-k2.6").teacher_base_url == opd_mod.RECIPE.opd.teacher_base_url
+    # unsupported teachers fail loudly on the worker (defensive guard, mirrors the kl_coef check).
+    for teacher in (
+        "gpt-5.5",
+        "deepseek-v4-pro",
+        "accounts/fireworks/models/deepseek-v4-pro",
+    ):
+        with pytest.raises(RuntimeError, match="teacher_model"):
+            _knobs(teacher)
 
 
 def test_groupwise_alignment_emits_no_empty_student_group():
@@ -2371,8 +2375,8 @@ class _TinyLM:
     "teacher_model",
     [
         "",
+        "glm-5.2",
         "accounts/fireworks/models/glm-5p2",
-        "accounts/fireworks/models/deepseek-v4-pro",
     ],
 )
 def test_opd_worker_rejects_nonvision_teacher_before_gpu_or_teacher_use(monkeypatch, teacher_model):

{freesolo_flash_dev-1.0.75 → freesolo_flash_dev-1.0.76}/uv.lock

@@ -1025,7 +1025,7 @@ wheels = [
 
 [[package]]
 name = "freesolo-flash"
-version = "1.0.75"
+version = "1.0.76"
 source = { editable = "." }
 
 [package.optional-dependencies]