freesolo-flash-dev 1.2.130__tar.gz → 1.2.132__tar.gz

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/auto-rebake.yml

@@ -232,7 +232,7 @@ jobs:
           sudo docker image prune -af || true
           df -h /
 
-      - uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
+      - uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4.4.1
 
       - name: Log in to GHCR
         uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
@@ -255,7 +255,7 @@ jobs:
           fi
 
       - name: Re-layer cache onto the fresh base (no GPU)
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7.4.0
         with:
           context: .
           file: docker/Dockerfile.kernelcache.relayer

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/bake-kernel-cache.yml

@@ -195,7 +195,7 @@ jobs:
 
       - name: Install uv
         if: steps.gate.outputs.run == 'true'
-        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"
@@ -237,7 +237,7 @@ jobs:
           sudo docker image prune -af || true
           df -h /
 
-      - uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
+      - uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4.4.1
         if: steps.gate.outputs.run == 'true'
 
       # thin layer FROM the SAME base the warmup ran inside -> the cache toolchain always matches the
@@ -245,7 +245,7 @@ jobs:
       # published base and bake a cache that silently won't load). GHCR auth from the login step above.
       - name: Build + push baked per-sm image (thin layer on the warmed base)
         if: steps.gate.outputs.run == 'true'
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7.4.0
         with:
           context: .
           file: docker/Dockerfile.kernelcache

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/ci.yml

@@ -33,7 +33,7 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install uv
-        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"
@@ -163,7 +163,7 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Install uv
-        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/deploy-modal-dev.yml

@@ -98,7 +98,7 @@ jobs:
         with:
           python-version: "3.12"
 
-      - uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+      - uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         if: steps.gate.outputs.deploy == 'true'
         with:
           # pin the installed binary as well as the action implementation.

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/publish-dev.yml

@@ -136,7 +136,7 @@ jobs:
 
       - name: Install uv
         if: steps.decide.outputs.publish == 'true'
-        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/publish-image.yml

@@ -43,7 +43,7 @@ jobs:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
       - name: Set up Docker Buildx
-        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
+        uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4.4.1
 
       - name: Log in to GHCR
         uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
@@ -98,7 +98,7 @@ jobs:
       # published here still runs on a plain --env-file. The Dockerfile defaults the argument
       # to false, so `docker build .` from a checkout stays vendor-free.
       - name: Build and push
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7.4.0
         with:
           context: .
           file: ./Dockerfile

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/publish-serving-image.yml

@@ -150,7 +150,7 @@ jobs:
 
       - name: Set up Docker Buildx
         if: steps.gate.outputs.publish == 'true'
-        uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
+        uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4.4.1
 
       - name: Log in to GHCR
         if: steps.gate.outputs.publish == 'true'
@@ -184,7 +184,7 @@ jobs:
       - name: Build and push
         if: steps.gate.outputs.publish == 'true'
         id: build
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7.4.0
         with:
           context: .
           file: ./Dockerfile.serve

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/publish.yml

@@ -202,7 +202,7 @@ jobs:
 
       - name: Install uv
         if: github.event_name == 'workflow_dispatch' || steps.changes.outputs.package_changed == 'true'
-        uses: astral-sh/setup-uv@20cfd1bf945f4377ade1205e4dbc17946fc9a30d # v10.0.1
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/.github/workflows/worker-image.yml

@@ -129,7 +129,7 @@ jobs:
           sudo docker image prune -af || true
           df -h /
 
-      - uses: docker/setup-buildx-action@37fe631027851001ddb9b187196cc803df7f5f0e # v4.3.0
+      - uses: docker/setup-buildx-action@f87e5991a6d7451dcb8d9637bfbc97413f497069 # v4.4.1
 
       - name: Log in to GHCR
         uses: docker/login-action@dbcb813823bdd20940b903addbd779551569679f # v4.6.0
@@ -140,7 +140,7 @@ jobs:
 
       - name: Build + push worker image
         id: image
-        uses: docker/build-push-action@53b7df96c91f9c12dcc8a07bcb9ccacbed38856a # v7.3.0
+        uses: docker/build-push-action@c3c9e263c25d99ce0380d002d59b67737d91b0dc # v7.4.0
         with:
           context: .
           file: Dockerfile.worker

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.2.130
+Version: 1.2.132
 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
@@ -29,7 +29,7 @@ Requires-Dist: pillow>=11; extra == 'dev'
 Requires-Dist: pytest>=9.0.3; extra == 'dev'
 Requires-Dist: referencing>=0.28.4; extra == 'dev'
 Requires-Dist: regex>=2024.11.6; extra == 'dev'
-Requires-Dist: ruff<0.17,>=0.6; extra == 'dev'
+Requires-Dist: ruff<0.18,>=0.6; extra == 'dev'
 Requires-Dist: runpod-flash<2,>=1.19; extra == 'dev'
 Requires-Dist: tiktoken>=0.12; extra == 'dev'
 Requires-Dist: tokenizers>=0.22; extra == 'dev'
@@ -63,7 +63,7 @@ Requires-Dist: freesolo>=0.4.2; extra == 'server'
 Requires-Dist: httpx>=0.27; extra == 'server'
 Requires-Dist: huggingface-hub>=1.2.0; extra == 'server'
 Requires-Dist: jsonschema>=4.23; extra == 'server'
-Requires-Dist: modal==1.5.4; extra == 'server'
+Requires-Dist: modal==1.6.1; extra == 'server'
 Requires-Dist: pillow>=11; extra == 'server'
 Requires-Dist: referencing>=0.28.4; extra == 'server'
 Requires-Dist: regex>=2024.11.6; extra == 'server'
@@ -73,7 +73,7 @@ Requires-Dist: uvicorn<1,>=0.52; extra == 'server'
 Provides-Extra: serving
 Requires-Dist: fastapi<1,>=0.136; extra == 'serving'
 Requires-Dist: httpx>=0.27.0; extra == 'serving'
-Requires-Dist: modal==1.5.4; extra == 'serving'
+Requires-Dist: modal==1.6.1; extra == 'serving'
 Requires-Dist: orjson>=3.10.0; extra == 'serving'
 Requires-Dist: pillow>=11; extra == 'serving'
 Requires-Dist: pydantic-settings>=2.6.0; extra == 'serving'

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/cli/ui/render.py

@@ -484,9 +484,9 @@ def _humanize_ts(value) -> str | None:
     """Format an epoch seconds value as a compact UTC timestamp, leaving non-numbers alone."""
     if not isinstance(value, (int, float)) or value <= 0:
         return None
-    import datetime
+    import datetime as dt
 
-    return datetime.datetime.fromtimestamp(value, datetime.UTC).strftime("%Y-%m-%d %H:%M UTC")
+    return dt.datetime.fromtimestamp(value, dt.UTC).strftime("%Y-%m-%d %H:%M UTC")
 
 
 def gpu_label(spec: dict, remote: dict) -> str:

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/core/spec.py

@@ -20,6 +20,7 @@ detail. See their docstrings before changing either.
 from __future__ import annotations
 
 import json
+import math
 import os
 from dataclasses import asdict, dataclass, field, fields
 from typing import Any, Literal
@@ -365,6 +366,9 @@ class TrainSpec:
     hf_repo: str = ""
     # None -> worker's tuned recipe default.
     learning_rate: float | None = field(default=None, metadata={"introduced_in": "0.2.0"})
+    # sft only: none preserves a legacy persisted run's 16x optimizer. new configs freeze the
+    # recipe default at parse time; 1 uses equal adapter learning rates.
+    loraplus_ratio: float | None = field(default=None, metadata={"introduced_in": "1.2.131"})
     # sft only. the packaged-dataset estimate resolves this against the selected row count into
     # examples_per_update (packed) or pins the optimizer batch to 1 (unpacked). grpo/opd have no
     # profile, so they take prompts_per_step instead and reject this key: the two are not the same
@@ -399,6 +403,17 @@ class TrainSpec:
     )
 
     def __post_init__(self) -> None:
+        if self.loraplus_ratio is not None:
+            ratio = self.loraplus_ratio
+            if isinstance(ratio, bool) or not isinstance(ratio, (int, float)):
+                raise TypeError("train.loraplus_ratio must be a number")
+            try:
+                ratio = float(ratio)
+            except OverflowError as exc:
+                raise ValueError("train.loraplus_ratio must be a finite number >= 1") from exc
+            if not math.isfinite(ratio) or ratio < 1:
+                raise ValueError("train.loraplus_ratio must be a finite number >= 1")
+            object.__setattr__(self, "loraplus_ratio", ratio)
         if not self.lora_alpha:
             object.__setattr__(self, "lora_alpha", 2 * self.lora_rank)
         max_steps = parse_max_steps(self.max_steps)
@@ -566,6 +581,8 @@ class JobSpec:
     project: str = ""
 
     def __post_init__(self) -> None:
+        if self.train.loraplus_ratio is not None and self.algorithm != "sft":
+            raise ValueError("train.loraplus_ratio only applies to sft")
         object.__setattr__(self, "seed", parse_seed(self.seed))
         object.__setattr__(self, "model_revision", _model_revision(self.model_revision))
         for field_name in ("model_revision_auto", "model_revision_force_pin"):
@@ -681,6 +698,9 @@ class JobSpec:
         differently, so removing either pop would break existing digests.
         """
         data = asdict(self)
+        # preserve historical preparation digests when the new option was not authored.
+        if data["train"]["loraplus_ratio"] is None:
+            data["train"].pop("loraplus_ratio")
         # missing and unset are the same internal state. omitting the empty value preserves that state
         # without serializing it into an explicit empty preference, which every parser rejects.
         if not data["gpu"].get("providers"):
@@ -786,6 +806,7 @@ class JobSpec:
                 init_from_adapter_revision=str(train.get("init_from_adapter_revision") or ""),
                 hf_repo=str(train.get("hf_repo") or ""),
                 learning_rate=opt_float(train.get("learning_rate")),
+                loraplus_ratio=train.get("loraplus_ratio"),
                 batch_size=opt_int(train.get("batch_size")),
                 prompts_per_step=opt_int(train.get("prompts_per_step")),
                 max_context_tokens=opt_int(train.get("max_context_tokens")),

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/engine/plan/recipe.py

@@ -170,6 +170,7 @@ class SFTConfig:
     max_seq_len: int = 1024
     max_seq_len_thinking: int = 2048
     learning_rate: float = 1e-4
+    loraplus_ratio: float = 1.0
     warmup_frac: float = 0.03
     effective_batch: int = 32
     num_epochs: int = 2

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/engine/worker/train/entry/sft_train.py

@@ -28,7 +28,6 @@ from flash.engine.worker.verl.checkpoints import restore_verl_resume
 from flash.engine.worker.verl.parallelism import ULYSSES_SEQUENCE_PARALLEL_SIZE
 
 # todo: run the two-gpu sft smoke on the exact runpod image and command assembled below.
-_SFT_LORAPLUS_RATIO = 16.0
 # consecutive zero-grad-norm steps tolerated before the run is failed as untrainable (GRAD-001).
 # any nonzero grad norm is proof the backward graph is intact and resets the count. 2 is enough to
 # separate a one-off fully-masked batch from a severed graph, and keeps the wasted spend to a couple
@@ -488,6 +487,7 @@ def _write_sft_result(options, data, model, child, progress, verified, outputs)
             "device_peak_gpu_gb": device_peak_gpu_gb,
             "loraplus_optim": _VERL_OPTIMIZER_NAME,
             "loraplus_applied": progress.loraplus_applied,
+            "loraplus_ratio": options.loraplus_ratio,
             "verl_backend": "fsdp2",
             # sft shards by DATA -- see ULYSSES_SEQUENCE_PARALLEL_SIZE for why. fsdp splits the batch
             # across the ranks actually LAUNCHED, which is the allocated card count only when the
@@ -585,7 +585,7 @@ def run_sft_train(spec=None) -> None:
     child = _prepare_sft_child(
         options, data, model, capabilities, use_remove_padding, gdn_reset_arch
     )
-    child_progress = _prepare_sft_progress(data, model, child)
+    child_progress = _prepare_sft_progress(options, data, model, child)
     progress = child_progress.values
 
     def on_line(line: str) -> None:

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/engine/worker/train/entry/sft_train_runner.py

@@ -33,7 +33,6 @@ from flash.providers.core.base import rentable_gpu_counts
 
 RECIPE = _sft_train.RECIPE
 _MAX_ZERO_GRAD_STEPS = _sft_train._MAX_ZERO_GRAD_STEPS
-_SFT_LORAPLUS_RATIO = _sft_train._SFT_LORAPLUS_RATIO
 _LORAPLUS_READY_MARKER = _sft_train._LORAPLUS_READY_MARKER
 _VERL_OPTIMIZER_IMPL = _sft_train._VERL_OPTIMIZER_IMPL
 _VERL_OPTIMIZER_NAME = _sft_train._VERL_OPTIMIZER_NAME
@@ -70,6 +69,7 @@ class _SftOptions:
     model_revision: str
     epochs: int
     learning_rate: float
+    loraplus_ratio: float
     effective_batch: int
     max_steps: int
     save_at_steps: tuple[int, ...]
@@ -147,6 +147,7 @@ class _SftProgress:
     loss_curve: list[float]
     train_tokens: int
     loraplus_applied: bool
+    loraplus_ratio: float
     wandb_link: dict[str, str | None]
     # what the child plugin must prove applied; verified at the first optimizer step and again at
     # child exit.
@@ -187,6 +188,9 @@ def _resolve_sft_options(spec) -> _SftOptions:
     model_id = spec.model if spec else RECIPE.hf_model_id
     model_revision = getattr(spec, "model_revision", "") if spec else ""
     train_spec = spec.train if spec else None
+    # new authored sft specs persist their ratio. an unset stored value predates the option and
+    # must retain the 16x optimizer groups used by its checkpoint, including across further retries.
+    default_loraplus_ratio = 16.0 if train_spec is not None else RECIPE.sft.loraplus_ratio
 
     def train_opt(name, default):
         value = getattr(train_spec, name, None) if train_spec else None
@@ -215,6 +219,7 @@ def _resolve_sft_options(spec) -> _SftOptions:
         model_revision=model_revision,
         epochs=int(train_opt("epochs", RECIPE.sft.num_epochs)),
         learning_rate=float(train_opt("learning_rate", RECIPE.sft.learning_rate)),
+        loraplus_ratio=float(train_opt("loraplus_ratio", default_loraplus_ratio)),
         effective_batch=int(train_opt("batch_size", RECIPE.sft.effective_batch)),
         max_steps=int(train_opt("max_steps", 0) or 0),
         save_at_steps=tuple(getattr(train_spec, "save_at_steps", ()) or ()),
@@ -588,7 +593,7 @@ def _write_sft_child_shims(
         {
             "marker_file": shim_markers,
             "seed": int(seed),
-            "loraplus_ratio": float(_SFT_LORAPLUS_RATIO),
+            "loraplus_ratio": options.loraplus_ratio,
             "loraplus_ready_marker": _LORAPLUS_READY_MARKER,
             "save_at_steps": list(options.save_at_steps),
             "total_steps": int(model.update_horizon),
@@ -758,7 +763,9 @@ def _prepare_sft_child(
     )
 
 
-def _prepare_sft_progress(data: _SftData, model: _SftModelSetup, child: _SftChild) -> _SftProgress:
+def _prepare_sft_progress(
+    options: _SftOptions, data: _SftData, model: _SftModelSetup, child: _SftChild
+) -> _SftProgress:
     resume_step = child.resume_step
     progress = {"step": resume_step, "loss": None, "grad_norm": None, "lr": None}
     # consecutive steps seen with grad_norm == 0.0 at a nonzero lr. one step can legitimately be
@@ -785,7 +792,8 @@ def _prepare_sft_progress(data: _SftData, model: _SftModelSetup, child: _SftChil
         observed_grad_norms=observed_grad_norms,
         loss_curve=[],
         train_tokens=train_tokens,
-        loraplus_applied=resume_step >= model.update_horizon,
+        loraplus_applied=options.loraplus_ratio > 1 and resume_step >= model.update_horizon,
+        loraplus_ratio=options.loraplus_ratio,
         wandb_link={},
         shim_markers=child.shim_markers,
         expected_shims=child.expected_shims,
@@ -846,7 +854,7 @@ def _consume_sft_marker_line(progress: _SftProgress, line: str) -> bool:
     if progress.expected_shims and not progress.shims_verified:
         verify_applied_shim_markers(progress.shim_markers, progress.expected_shims)
         progress.shims_verified = True
-    if _SFT_LORAPLUS_RATIO > 1 and not progress.loraplus_applied:
+    if progress.loraplus_ratio > 1 and not progress.loraplus_applied:
         raise RuntimeError(
             "verl reached an optimizer step before the required lora+ shim succeeded"
         )
@@ -914,7 +922,7 @@ def _finish_sft_child(
     if progress.expected_shims and not progress.shims_verified:
         verify_applied_shim_markers(progress.shim_markers, progress.expected_shims)
         progress.shims_verified = True
-    if _SFT_LORAPLUS_RATIO > 1 and not progress.loraplus_applied:
+    if progress.loraplus_ratio > 1 and not progress.loraplus_applied:
         raise RuntimeError("required lora+ shim did not emit its success marker")
     return train_wall, device_peak_gpu_gb
 

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/schema/__init__.py

@@ -25,6 +25,7 @@ from flash.core.spec import (
     parse_seed,
     require_project_id,
 )
+from flash.engine.plan.recipe import RECIPE
 from flash.providers.core.base import (
     UnsupportedGpuError,
     authored_gpu_ceiling,
@@ -297,6 +298,8 @@ def validate_train_keys_for_algorithm(train_raw: Mapping[str, Any], algorithm: s
     every spec fail to reparse its own output and break resubmit, warm-start and server reparse.
     A null is the absence of an authored value, which is exactly what the user is being asked for.
     """
+    if train_raw.get("loraplus_ratio") is not None and algorithm != "sft":
+        raise ConfigError("train.loraplus_ratio only applies to sft")
     for key, allowed in _ALGORITHM_ONLY_TRAIN_KEYS.items():
         if train_raw.get(key) is None or algorithm in allowed:
             continue
@@ -678,6 +681,10 @@ def spec_from_dict(
     )
     wandb_spec = _validate_wandb_section(raw)
 
+    loraplus_ratio = _train_float(train_raw, "loraplus_ratio", minimum=1.0)
+    if algorithm == "sft" and loraplus_ratio is None:
+        # freeze the new default; an absent persisted value identifies the legacy 16x optimizer.
+        loraplus_ratio = RECIPE.sft.loraplus_ratio
     try:
         train_spec = TrainSpec(
             epochs=_train_int(train_raw, "epochs", minimum=1),
@@ -686,6 +693,7 @@ def spec_from_dict(
             init_from_adapter=init_from_adapter,
             hf_repo="",  # assigned server-side; see submit_job._assign_managed_hf_repo
             learning_rate=_train_float(train_raw, "learning_rate", minimum=0.0, exclusive=True),
+            loraplus_ratio=loraplus_ratio,
             batch_size=_train_int(train_raw, "batch_size", minimum=1),
             prompts_per_step=_train_int(train_raw, "prompts_per_step", minimum=1),
             max_context_tokens=_train_int(train_raw, "max_context_tokens", minimum=1),

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/schema/fields.py

@@ -65,7 +65,10 @@ def _train_float(
         return None
     if isinstance(v, bool) or not isinstance(v, (int, float)):
         raise ConfigError(f"train.{key} must be a number")
-    v = float(v)
+    try:
+        v = float(v)
+    except OverflowError as exc:
+        raise ConfigError(f"train.{key} must be a finite number") from exc
     # nan/inf slip past range checks (nan compares false, inf passes any minimum).
     if not math.isfinite(v):
         raise ConfigError(f"train.{key} must be a finite number")

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/flash/serve/runtime/sampling.py

@@ -121,6 +121,9 @@ def _normalize_token_position(
             selected = record
     if selected is None:
         raise RuntimeNotReadyError("vllm logprobs omitted the selected token")
+    # vllm may insert the sampled token before higher-probability alternatives.
+    # rank before truncating; token id makes equal-probability ties deterministic.
+    normalized.sort(key=lambda item: (-item[1]["logprob"], item[0]))
     return {
         **selected,
         "top_logprobs": [record for _, record in normalized[:top_logprobs]],

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "freesolo-flash-dev"
-version = "1.2.130"
+version = "1.2.132"
 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).
@@ -60,7 +60,7 @@ server = [
     "fastapi>=0.136,<1",
     "uvicorn>=0.52,<1",
     # customer-owned modal lifecycle mutations execute in the control plane.
-    "modal==1.5.4",
+    "modal==1.6.1",
     "runpod-flash>=1.19,<2",
     "httpx>=0.27",
     "jsonschema>=4.23",
@@ -98,7 +98,7 @@ serving = [
     # this one is what the serving image installs ALONE -- so a range here lets the container
     # resolve a different modal than the control plane was validated against, on a rebuild that
     # changed nothing else. one policy per package.
-    "modal==1.5.4",
+    "modal==1.6.1",
     "orjson>=3.10.0",
     # `src/multimodal.py` decodes and normalizes inbound images at module scope. the `gpu` extra
     # gets Pillow transitively via vllm, but this extra deliberately excludes vllm, so it has to
@@ -134,11 +134,11 @@ dev = [
     "pytest>=9.0.3",
     # dependency contract tests parse exact pins without relying on a transitive install.
     "packaging>=24,<27",
-    # Upper bound tracks uv.lock (ruff 0.16.x), which is what CI lints with. Without it, an
+    # Upper bound tracks uv.lock (ruff 0.17.x), which is what CI lints with. Without it, an
     # install that ignores the lockfile (`uv pip install -e '.[dev]'`, plain pip) resolves whatever
     # ruff is newest, whose added rules fail on a tree CI calls clean. Raise both sites together
     # with the lockfile, never ahead of it.
-    "ruff>=0.6,<0.17",
+    "ruff>=0.6,<0.18",
     # mirror server's 0.136 floor so dev remains compatible with the deployed serving extra pair.
     "fastapi>=0.136,<1",
     "uvicorn>=0.52,<1",
@@ -196,7 +196,7 @@ 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 = "1.2.130"
+version = "1.2.132"
 
 [tool.pytest.ini_options]
 pythonpath = ["."]
@@ -232,7 +232,7 @@ dev = [
     # dependency contract tests parse exact pins without relying on a transitive install.
     "packaging>=24,<27",
     # Keep in lockstep with the `dev` extra above and with uv.lock.
-    "ruff>=0.6,<0.17",
+    "ruff>=0.6,<0.18",
     # Offline worker/env tests import these without the heavy `gpu` stack; the CI test env
     # (`--extra server --dev`) gets httpx + huggingface-hub from the `server` extra.
     "datasets>=2.19",

{freesolo_flash_dev-1.2.130 → freesolo_flash_dev-1.2.132}/tests/test_lora_targeting.py

@@ -456,7 +456,7 @@ def test_sft_grpo_and_opd_plugins_install_text_targeting_only_for_text_jobs(tmp_
     shim_dir = tmp_path / "sft-text-shim"
     shim_dir.mkdir()
     _marker, expected, raw_sft = sft_train_runner._write_sft_child_shims(
-        SimpleNamespace(model_id="Qwen/Qwen3.5-9B", save_at_steps=()),
+        SimpleNamespace(model_id="Qwen/Qwen3.5-9B", save_at_steps=(), loraplus_ratio=16.0),
         SimpleNamespace(
             update_horizon=1,
             reentrant_gradient_checkpointing=False,
@@ -474,7 +474,7 @@ def test_sft_grpo_and_opd_plugins_install_text_targeting_only_for_text_jobs(tmp_
     multimodal_dir = tmp_path / "sft-multimodal-shim"
     multimodal_dir.mkdir()
     _marker, expected, raw_sft = sft_train_runner._write_sft_child_shims(
-        SimpleNamespace(model_id="Qwen/Qwen3.5-9B", save_at_steps=()),
+        SimpleNamespace(model_id="Qwen/Qwen3.5-9B", save_at_steps=(), loraplus_ratio=16.0),
         SimpleNamespace(
             update_horizon=1,
             reentrant_gradient_checkpointing=False,

freesolo_flash_dev-1.2.132/tests/test_loraplus_config.py

@@ -0,0 +1,97 @@
+"""user-authored lora+ ratios survive cli and worker spec boundaries."""
+
+import pytest
+
+from flash.core.spec import JobSpec, TrainSpec
+from flash.schema import ConfigError, spec_and_train_keys_from_file, spec_from_dict
+
+
+def _raw(train, algorithm="sft"):
+    return {
+        "model": "Qwen/Qwen3.5-9B",
+        "algorithm": algorithm,
+        "project": "11111111-1111-4111-8111-111111111111",
+        "environment": {"id": "owner/project/env"},
+        "train": train,
+    }
+
+
+@pytest.mark.parametrize("ratio", [1, 3.5, 16, 32])
+def test_loraplus_ratio_survives_public_and_worker_round_trips(ratio):
+    spec = spec_from_dict(_raw({"loraplus_ratio": ratio, "learning_rate": 1e-5}))
+    public = spec_from_dict(spec.to_dict())
+    worker = JobSpec.from_json(spec.to_json())
+    for result in (spec, public, worker):
+        assert result.train.loraplus_ratio == ratio
+        assert result.train.learning_rate == 1e-5
+
+
+@pytest.mark.parametrize("train", [{}, {"loraplus_ratio": None}])
+def test_new_unset_ratio_is_frozen_for_future_retries(train):
+    spec = spec_from_dict(_raw(train))
+    assert spec.train.loraplus_ratio == 1.0
+    assert spec.to_dict()["train"]["loraplus_ratio"] == 1.0
+    assert JobSpec.from_json(spec.to_json()).train.loraplus_ratio == 1.0
+
+
+def test_legacy_unset_ratio_preserves_historical_payload_shape():
+    spec = JobSpec.from_dict(_raw({}))
+    assert spec.train.loraplus_ratio is None
+    assert "loraplus_ratio" not in spec.to_dict()["train"]
+    assert "loraplus_ratio" not in spec.to_internal_dict()["train"]
+    assert JobSpec.from_json(spec.to_json()).train.loraplus_ratio is None
+
+
+@pytest.mark.parametrize(
+    "ratio",
+    [
+        True,
+        False,
+        "16",
+        [],
+        0,
+        -1,
+        0.5,
+        float("nan"),
+        float("inf"),
+        -float("inf"),
+        10**400,
+        -(10**400),
+    ],
+)
+def test_invalid_ratio_is_rejected_at_public_and_internal_boundaries(ratio):
+    with pytest.raises(ConfigError, match="loraplus_ratio"):
+        spec_from_dict(_raw({"loraplus_ratio": ratio}))
+    with pytest.raises((TypeError, ValueError), match="loraplus_ratio"):
+        JobSpec.from_dict(_raw({"loraplus_ratio": ratio}))
+    with pytest.raises((TypeError, ValueError), match="loraplus_ratio"):
+        TrainSpec(loraplus_ratio=ratio)
+
+
+@pytest.mark.parametrize("algorithm", ["grpo", "opd"])
+def test_ratio_is_rejected_for_algorithms_without_loraplus(algorithm):
+    raw = _raw({"loraplus_ratio": 4}, algorithm)
+    with pytest.raises(ConfigError, match="only applies to sft"):
+        spec_from_dict(raw)
+    with pytest.raises(ValueError, match="only applies to sft"):
+        JobSpec.from_dict(raw)
+
+
+def test_toml_ratio_and_cli_override_are_user_authored_keys(tmp_path):
+    config = tmp_path / "flash.toml"
+    config.write_text(
+        'model = "Qwen/Qwen3.5-9B"\n'
+        'algorithm = "sft"\n'
+        'project = "11111111-1111-4111-8111-111111111111"\n'
+        '[environment]\nid = "owner/project/env"\n'
+        "[train]\nloraplus_ratio = 4.0\nlearning_rate = 0.00001\n"
+    )
+    spec, keys = spec_and_train_keys_from_file(str(config), run_id="test")
+    assert spec.train.loraplus_ratio == 4.0
+    assert "loraplus_ratio" in keys
+    spec, keys = spec_and_train_keys_from_file(
+        str(config), run_id="test", overrides=["train.loraplus_ratio=1.0"]
+    )
+    assert spec.train.loraplus_ratio == 1.0
+    assert spec.train.learning_rate == 1e-5
+    assert "loraplus_ratio" in keys