freesolo-flash-dev 1.0.77__tar.gz → 1.0.78__tar.gz

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/Dockerfile.worker

@@ -268,7 +268,7 @@ print('worker stack OK | torch', torch.__version__, '| cudnn', cudnn, '| triton'
 # flash's, so a job that sets WANDB_API_KEY loses its curves and run link unless wandb is importable
 # in here; resolve_verl_loggers() degrades to console when it is not, which is the right behaviour
 # for a foreign FLASH_VERL_PYTHON but silently drops W&B on the default image (codex[bot]).
-ARG VERL_SPEC=verl[vllm]@git+https://github.com/freesolo-co/verl@b7492fa3b7ab843294d06dbf754e887950f559c7
+ARG VERL_SPEC=verl[vllm]@git+https://github.com/freesolo-co/verl@1bea7d6825bbb9d2164e86e379b3680e7c53bb8a
 ENV FLASH_VERL_PYTHON=/opt/verl-venv/bin/python
 RUN printf '%s\n' "numpy==2.2.6" "xgrammar==0.1.25" "vllm==0.19.1" > /tmp/verl-overrides.txt \
     && uv venv --seed /opt/verl-venv \

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: freesolo-flash-dev
-Version: 1.0.77
+Version: 1.0.78
 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-1.0.77 → freesolo_flash_dev-1.0.78}/flash/catalog.py

@@ -369,6 +369,7 @@ MODELS: dict[str, ModelInfo] = {
         ),
         algos=ALGORITHMS,
         min_vram_gb=48,
+        min_disk_gb=120,
         # NOT QLoRA: peft bnb merge during GRPO rollout diverges trainer precision -> TRL ratio collapses to 0.
         grpo_min_vram_gb=80,
         quant="bf16",

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/flash/engine/worker/backend_common.py

@@ -29,13 +29,13 @@ from http.server import ThreadingHTTPServer
 
 from flash.diagnostics import sanitize_diagnostic
 
-# verl 0.8.0 exactly, plus the truncation-mask and 3d position id commits. it must stay on the 0.8.0
-# base: the opd plugin patches 0.8.0 internals and imports verl.trainer.main_ppo_sync, which verl
-# deleted after 0.8.0, and opd's exact-version gate reads the version file this branch pins to the
-# release value.
+# verl 0.8.0 exactly, plus the truncation-mask, 3d position-id, and ulysses fused-label fixes.
+# it must stay on the 0.8.0 base: the opd plugin patches 0.8.0 internals and imports
+# verl.trainer.main_ppo_sync, which verl deleted after 0.8.0, and opd's exact-version gate reads the
+# version file this branch pins to the release value.
 VERL_REQUIREMENT_NAME = "verl"
 VERL_REQUIREMENT_URL = (
-    "git+https://github.com/freesolo-co/verl@ee327f43c87c031034ddb88c17fefe03d20a0505"
+    "git+https://github.com/freesolo-co/verl@1bea7d6825bbb9d2164e86e379b3680e7c53bb8a"
 )
 # the pin, as the venv stamp records it. the provisioning install asks for the [vllm] extra of this
 # same commit; the stamp stays extra-free so it identifies the verl a venv holds, not how it was

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.0.77"
+version = "1.0.78"
 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 = "1.0.77"
+version = "1.0.78"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/tests/test_backend_common.py

@@ -371,7 +371,7 @@ def test_resolve_verl_python_installs_pinned_gpu_dependencies(monkeypatch, tmp_p
     assert "cp312" in vc.FLASH_ATTN_SPEC
     install = calls[1]
     assert vc.VERL_REQUIREMENT == (
-        "verl @ git+https://github.com/freesolo-co/verl@ee327f43c87c031034ddb88c17fefe03d20a0505"
+        "verl @ git+https://github.com/freesolo-co/verl@1bea7d6825bbb9d2164e86e379b3680e7c53bb8a"
     )
     assert any(vc.VERL_REQUIREMENT_URL in arg for arg in install)
     assert "liger-kernel" in install
@@ -548,6 +548,16 @@ def test_flash_attn_spec_stays_in_lockstep_with_the_worker_image():
     )
 
 
+def test_verl_spec_stays_in_lockstep_with_the_worker_image():
+    # the default image exports FLASH_VERL_PYTHON, so its baked interpreter bypasses the fallback
+    # resolver entirely. both paths must pin the same commit or live workers miss dependency fixes.
+    dockerfile = pathlib.Path(__file__).resolve().parents[1] / "Dockerfile.worker"
+    expected = f"ARG VERL_SPEC=verl[vllm]@{vc.VERL_REQUIREMENT_URL}"
+    assert expected in dockerfile.read_text(), (
+        "Dockerfile.worker's VERL_SPEC drifted from backend_common.VERL_REQUIREMENT_URL"
+    )
+
+
 def test_the_venv_stamp_records_the_pin_not_the_install_extras(monkeypatch, tmp_path):
     # the stamp gates rebuilds. if it recorded the extra-bearing spec while the pin stayed bare,
     # every later call would see a mismatch and rebuild the venv from scratch on a paid pod.
@@ -652,14 +662,14 @@ def test_verl_pin_matches_the_version_opd_requires_exactly():
 
     assert plugin._STRUCTURED_RUNTIME_EXACT_VERSIONS["verl"] == "0.8.0"
     # asserting the constant alone would let a newer-base commit land silently, so bind the pinned
-    # commit itself to that version. this is the sha of the truncation-mask and 3d position id
-    # commits cherry-picked onto the v0.8.0 tag, plus the agent-loop position-id pad fix; moving the
-    # pin must be a deliberate edit here, with the base re-verified. re-verified at ee327f43:
-    # verl/version/version still reads 0.8.0, verl/trainer/main_ppo_sync.py is still present, and the
-    # only diff against the previous pin is agent_loop.py.
+    # commit itself to that version. this is the sha of the truncation-mask and 3d position-id
+    # commits cherry-picked onto the v0.8.0 tag, plus the agent-loop position-id pad, fused-label
+    # shape fix, and qwen3.5 shift-label ownership fix. moving the pin must be deliberate, with the
+    # base re-verified. at 1bea7d68, verl/version/version still reads 0.8.0 and
+    # verl/trainer/main_ppo_sync.py is still present.
     _, _, ref = vc.VERL_REQUIREMENT.partition("git+")
     _, _, commit = ref.rpartition("@")
-    assert commit == "ee327f43c87c031034ddb88c17fefe03d20a0505"
+    assert commit == "1bea7d6825bbb9d2164e86e379b3680e7c53bb8a"
 
 
 def test_resolve_verl_python_installs_wandb_best_effort_when_requested(monkeypatch, tmp_path):

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/tests/test_disk_gb.py

@@ -48,8 +48,13 @@ def test_oversized_catalog_models_carry_disk_floors():
     """Models whose transient download peak exceeds the shared cache need per-job disk floors."""
     from flash.catalog import MODELS
 
-    disk_floor_models = {"Qwen/Qwen3.6-27B", "Qwen/Qwen3.6-35B-A3B"}
+    disk_floor_models = {
+        "Qwen/Qwen3.5-9B",
+        "Qwen/Qwen3.6-27B",
+        "Qwen/Qwen3.6-35B-A3B",
+    }
     assert all(m.min_disk_gb == 0 for m in MODELS.values() if m.id not in disk_floor_models)
+    assert MODELS["Qwen/Qwen3.5-9B"].min_disk_gb == 120
     assert MODELS["Qwen/Qwen3.6-27B"].min_disk_gb == 160
     assert MODELS["Qwen/Qwen3.6-35B-A3B"].min_disk_gb == 200
 

{freesolo_flash_dev-1.0.77 → freesolo_flash_dev-1.0.78}/uv.lock

@@ -1025,7 +1025,7 @@ wheels = [
 
 [[package]]
 name = "freesolo-flash"
-version = "1.0.77"
+version = "1.0.78"
 source = { editable = "." }
 
 [package.optional-dependencies]