freesolo-flash-dev 1.2.133__tar.gz → 1.2.134__tar.gz

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.2.133
+Version: 1.2.134
 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-1.2.133 → freesolo_flash_dev-1.2.134}/flash/core/spec.py

@@ -20,7 +20,6 @@ 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
@@ -366,9 +365,6 @@ 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
@@ -403,17 +399,6 @@ 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)
@@ -581,8 +566,6 @@ 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"):
@@ -698,9 +681,6 @@ 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"):
@@ -806,7 +786,6 @@ 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-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/plan/recipe.py

@@ -170,7 +170,6 @@ 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-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/support/verl_checkpoint.py

@@ -8,7 +8,6 @@ from collections.abc import Iterable
 from dataclasses import dataclass
 
 from flash._internal.fileio import reject_duplicate_keys
-from flash.engine.support.verl_policy import FsdpGeneration
 
 VERL_FSDP_CONFIG_FILE = "fsdp_config.json"
 _MAX_NATIVE_FSDP_WORLD_SIZE = 8
@@ -78,7 +77,6 @@ def inspect_fsdp_checkpoint_manifest(
     stamp_raw: bytes | str | None,
     files: Iterable[FsdpManifestFile],
     *,
-    expected_generation: FsdpGeneration,
     expected_world_size: int | None,
 ) -> FsdpCheckpointInspection:
     """validate the exact canonical native shard set for one stamped actor directory."""
@@ -90,13 +88,11 @@ def inspect_fsdp_checkpoint_manifest(
             rejection_reason="invalid or missing fsdp stamp",
         )
     generation, width = stamp
-    if generation != expected_generation:
+    if generation != 2:
         return FsdpCheckpointInspection(
             loadable=False,
             width=width,
-            rejection_reason=(
-                f"fsdp generation mismatch (stamped {generation}, expected {expected_generation})"
-            ),
+            rejection_reason=f"fsdp generation mismatch (stamped {generation}, expected 2)",
         )
     if expected_world_size is not None and width != expected_world_size:
         return FsdpCheckpointInspection(

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/entry/opd_train_runner.py

@@ -28,7 +28,6 @@ from flash.engine.plan import steps as _steps
 from flash.engine.plan.recipe import RECIPE
 from flash.engine.plan.steps import rl_data_parallel_cards
 from flash.engine.profiling.sft_workload import _materialize_verl_images
-from flash.engine.support.verl_policy import _resolve_fsdp_generation
 from flash.engine.worker.entry import opd as _opd_entry
 from flash.engine.worker.train.core.child.runtime import TEXT_LORA_TARGET_SHIM
 from flash.engine.worker.train.entry.prompt_rows import canonical_prompt_messages
@@ -187,7 +186,6 @@ def _prepare_workload(
     if isinstance(target_modules, set | frozenset):
         target_modules = sorted(target_modules)
     target_parameters = tuple(targeting.target_parameters) if targeting.target_parameters else None
-    fsdp_generation = _resolve_fsdp_generation("opd", target_parameters)
     lora_rank = int(lora_config.r)
     return _WorkloadState(
         prompts_per_step,
@@ -209,7 +207,6 @@ def _prepare_workload(
         int(lora_config.lora_alpha),
         target_modules,
         target_parameters,
-        fsdp_generation,
         targeting.exclude_modules,
         _sft._warmstart_adapter_path(
             request.model_id,
@@ -260,8 +257,6 @@ def _materialize_child_files(
         # the same count this attempt hands verl as n_gpus_per_node, which is the data-parallel
         # width: ulysses is pinned to 1, so every rank is a dp rank.
         world_size=gpu_count,
-        # native state is generation-specific, so use the same resolved policy as the child config.
-        expected_fsdp_generation=workload.fsdp_generation,
     )
     bridge = _opd_bridge._TeacherAlignmentBridge(
         prompts=prompt_state.prompts,
@@ -399,7 +394,6 @@ def _build_base_config(
         "target_modules": workload.target_modules,
         "exclude_modules": None,
         "target_parameters": workload.target_parameters,
-        "fsdp_generation": workload.fsdp_generation,
         "lora_adapter_path": workload.warmstart_adapter,
         "learning_rate": knobs.learning_rate,
         "local_dir": workload.local_dir,

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/entry/rl_train_runner.py

@@ -309,7 +309,6 @@ def _prepare_rl_files(inp, prompts):
     resume_step = rl_checkpoints._restore_verl_resume(
         local_dir,
         world_size=int(inp["dp_cards"]),
-        expected_fsdp_generation=inp["fsdp_generation"],
         required_steps=tuple(inp["save_at_steps"]),
     )
     train_pq = os.path.join(workdir, "train.parquet")

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/entry/sft_train.py

@@ -355,8 +355,7 @@ def _sft_liger_config() -> dict[str, bool]:
 
 
 # re-exported so the child-render helpers stay reachable as `sft_train.<name>`: the other
-# trainers import `_hydra_val` from here, and the sft tests import the renderers and the
-# LoRA+ marker from here.
+# trainers import `_hydra_val` from here, and the sft tests import the renderers.
 # the runner resolves patched seams through this parent module at call time, so these imports must
 # remain attributes of `sft_train` rather than direct imports in the split module.
 from flash.engine.plan.recipe import RECIPE  # noqa: E402,F401
@@ -394,7 +393,6 @@ from flash.engine.worker.train.sft.setup.checkpoints import (  # noqa: E402,F401
     _VerlCheckpointWatcher,
 )
 from flash.engine.worker.train.sft.setup.config import (  # noqa: E402,F401
-    _LORAPLUS_READY_MARKER,
     _REQUIRED_OVERRIDE_KEYS,
     _VERL_OPTIMIZER_IMPL,
     _VERL_OPTIMIZER_NAME,
@@ -485,9 +483,6 @@ def _write_sft_result(options, data, model, child, progress, verified, outputs)
             "loss_curve": progress.loss_curve[:400],
             "peak_gpu_gb": device_peak_gpu_gb,
             "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 +580,7 @@ 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(options, data, model, child)
+    child_progress = _prepare_sft_progress(data, model, child)
     progress = child_progress.values
 
     def on_line(line: str) -> None:
@@ -606,7 +601,7 @@ def run_sft_train(spec=None) -> None:
     if child.resume_step < model.update_horizon:
         child.watcher.start()
         # check save completeness only after normal training completion. callback failures occur before
-        # return_code assignment; checking anyway would replace the real zero-grad or lora+ diagnosis
+        # return_code assignment; checking anyway would replace the real zero-grad or runtime-patch diagnosis
         # with a missing-save error. opd_train uses the same guard.
         training_completed = False
         try:

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/entry/sft_train_runner.py

@@ -25,7 +25,6 @@ from flash.adapters.fused_experts import lora_target_parameters
 from flash.adapters.targets import resolve_lora_targeting
 from flash.core.catalog import get_model
 from flash.engine.plan.steps import sft_data_parallel_cards, widest_usable_dp_width
-from flash.engine.support.verl_policy import _resolve_fsdp_generation
 from flash.engine.worker.train.core.child.runtime import TEXT_LORA_TARGET_SHIM
 from flash.engine.worker.train.entry import sft_train as _sft_train
 from flash.engine.worker.verl.parallelism import ULYSSES_SEQUENCE_PARALLEL_SIZE
@@ -33,7 +32,6 @@ from flash.providers.core.base import rentable_gpu_counts
 
 RECIPE = _sft_train.RECIPE
 _MAX_ZERO_GRAD_STEPS = _sft_train._MAX_ZERO_GRAD_STEPS
-_LORAPLUS_READY_MARKER = _sft_train._LORAPLUS_READY_MARKER
 _VERL_OPTIMIZER_IMPL = _sft_train._VERL_OPTIMIZER_IMPL
 _VERL_OPTIMIZER_NAME = _sft_train._VERL_OPTIMIZER_NAME
 build_sft_overrides = _sft_train.build_sft_overrides
@@ -69,7 +67,6 @@ class _SftOptions:
     model_revision: str
     epochs: int
     learning_rate: float
-    loraplus_ratio: float
     effective_batch: int
     max_steps: int
     save_at_steps: tuple[int, ...]
@@ -146,8 +143,6 @@ class _SftProgress:
     observed_grad_norms: list[float]
     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.
@@ -188,9 +183,6 @@ 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
@@ -219,7 +211,6 @@ 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 ()),
@@ -593,8 +584,6 @@ def _write_sft_child_shims(
         {
             "marker_file": shim_markers,
             "seed": int(seed),
-            "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),
             "reentrant_gradient_checkpointing": bool(model.reentrant_gradient_checkpointing),
@@ -638,7 +627,6 @@ def _prepare_sft_child(
     # take the RAISING gate there rather than the soft form.
     world_size, micro_batch = _resolve_sft_width_and_micro_batch(options, data, model)
     target_parameters = lora_target_parameters(options.model_id)
-    fsdp_generation = _resolve_fsdp_generation("sft", target_parameters)
     config = {
         "train_files": data.train_file,
         "train_batch_size": model.train_batch_size,
@@ -652,7 +640,6 @@ def _prepare_sft_child(
         "target_modules": model.target_modules,
         "exclude_modules": None,
         "target_parameters": target_parameters,
-        "fsdp_generation": fsdp_generation,
         "lora_adapter_path": model.warmstart_adapter,
         "ulysses_sp_size": ULYSSES_SEQUENCE_PARALLEL_SIZE,
         "lr": options.learning_rate,
@@ -706,7 +693,6 @@ def _prepare_sft_child(
     resume_step = _sft_train._restore_verl_resume(
         options.paths.local_dir,
         world_size=world_size,
-        expected_fsdp_generation=fsdp_generation,
     )
     watcher = _sft_train._VerlCheckpointWatcher(
         local_dir=options.paths.local_dir,
@@ -763,9 +749,7 @@ def _prepare_sft_child(
     )
 
 
-def _prepare_sft_progress(
-    options: _SftOptions, data: _SftData, model: _SftModelSetup, child: _SftChild
-) -> _SftProgress:
+def _prepare_sft_progress(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
@@ -792,12 +776,10 @@ def _prepare_sft_progress(
         observed_grad_norms=observed_grad_norms,
         loss_curve=[],
         train_tokens=train_tokens,
-        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,
-        # like loraplus_applied above: a resume already at the horizon never launches the child,
+        # a resume already at the horizon never launches the child,
         # so there is no marker file to verify and nothing left for the shims to patch.
         shims_verified=resume_step >= model.update_horizon,
     )
@@ -838,26 +820,17 @@ def _invoke_sft_child(child: _SftChild, callbacks: _SftProgressCallbacks, on_lin
 def _consume_sft_marker_line(progress: _SftProgress, line: str) -> bool:
     """Fold a child log line's non-metric markers in; answer whether it is an optimizer step.
 
-    False means the caller has nothing further to do with the line. A True answer additionally
-    asserts the lora+ shim landed before the first step, because a run that reaches an optimizer
-    update without it is training plain lora at the lora+ learning rate.
+    False means the caller has nothing further to do with the line. Before the first optimizer
+    step, verify that the required child runtime patches applied.
     """
-    if _LORAPLUS_READY_MARKER in line:
-        progress.loraplus_applied = True
     link = _sft_train.parse_wandb_link(line)
     if link is not None:
         progress.wandb_link.update(link)
     if _sft_train.verl_step_number(line) is None:
         return False
-    # the marker set first: a skipped sitecustomize also loses the lora+ shim, and "no fragment
-    # ran at all" is the root cause worth reporting over its lora+ symptom.
     if progress.expected_shims and not progress.shims_verified:
         verify_applied_shim_markers(progress.shim_markers, progress.expected_shims)
         progress.shims_verified = True
-    if progress.loraplus_ratio > 1 and not progress.loraplus_applied:
-        raise RuntimeError(
-            "verl reached an optimizer step before the required lora+ shim succeeded"
-        )
     return True
 
 
@@ -918,12 +891,10 @@ def _finish_sft_child(
         raise RuntimeError(f"verl SFT subprocess exited with status {return_code}")
     # belt and braces behind the first-step check in _consume_sft_marker_line, for a child whose
     # step lines the parent never parsed. shims_verified starts True when the child is not
-    # launched at all (resume already at the horizon), exactly like loraplus_applied above.
+    # launched at all (resume already at the horizon).
     if progress.expected_shims and not progress.shims_verified:
         verify_applied_shim_markers(progress.shim_markers, progress.expected_shims)
         progress.shims_verified = True
-    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-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/opd/orchestration/failures.py

@@ -21,7 +21,6 @@ from pathlib import Path
 import flash.engine.worker.io.hf as _worker_hf
 import flash.engine.worker.perf as _worker_perf
 import flash.engine.worker.runtime.state as _worker_state
-from flash.engine.support.verl_policy import FsdpGeneration
 from flash.engine.worker.train.entry.sft_train import (
     SHIM_FRAGMENT_FAILED_EXIT_CODE,
     _export_checkpoint_adapter,
@@ -378,7 +377,6 @@ def _restore_verl_resume(
     prompt_pool_fingerprint: str,
     update_horizon: int,
     world_size: int,
-    expected_fsdp_generation: FsdpGeneration,
 ) -> tuple[int, dict | None]:
     revision = _worker_state.OPD_RESUME_REVISION or None
     # no `prefer`: opd pins an exact commit via OPD_RESUME_REVISION with fail_closed, and
@@ -402,7 +400,6 @@ def _restore_verl_resume(
         inspection = inspect_resume_checkpoint(
             resume,
             world_size=world_size,
-            expected_fsdp_generation=expected_fsdp_generation,
         )
         if not inspection.loadable:
             raise RuntimeError(
@@ -410,12 +407,11 @@ def _restore_verl_resume(
                 f"checkpoint {os.path.basename(resume)}, rejected because "
                 f"{inspection.diagnostic()}. restarting from step 0 would violate the pinned-resume "
                 "contract, so this attempt refuses to train; relaunch from a compatible complete "
-                f"fsdp{expected_fsdp_generation} checkpoint."
+                "fsdp2 checkpoint."
             )
     elif not resume_topology_matches(
         resume,
         world_size=world_size,
-        expected_fsdp_generation=expected_fsdp_generation,
         job_label="OPD",
     ):
         return 0, None

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/opd/orchestration/overrides.py

@@ -39,7 +39,6 @@ _REQUIRED_OVERRIDE_KEYS = (
     "lora_rank",
     "lora_alpha",
     "target_modules",
-    "fsdp_generation",
     "learning_rate",
     "local_dir",
     "save_freq",
@@ -150,9 +149,9 @@ def _actor_rollout_overrides(config: dict, *, max_tokens: int) -> list[str]:
         # absent -> True, verl's own default: a config built without the gate must render the
         # lower-memory strategy, never fall into zero-2 by omission.
         f"actor_rollout_ref.actor.fsdp_config.reshard_after_forward={_hydra_val(bool(config.get('reshard_after_forward', True)))}",
-        # dense opd stays on the live-validated fsdp1 path. fused-expert target_parameters requires
-        # fsdp2 so peft can parametrize named tensors. reshard_after_forward remains independent.
-        *actor_fsdp_strategy_overrides(config["fsdp_generation"]),
+        # dense and fused-expert opd both use fsdp2, matching sft and grpo. the independent
+        # reshard_after_forward setting still selects zero-2 or zero-3.
+        *actor_fsdp_strategy_overrides(),
         # store the frozen base in bf16, not verl's fp32 yaml default. shared with the rl driver.
         *trainer_dtype_overrides(),
         "actor_rollout_ref.rollout.name=vllm",

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/opd/orchestration/state.py

@@ -5,8 +5,6 @@ from __future__ import annotations
 from dataclasses import dataclass
 from typing import Any
 
-from flash.engine.support.verl_policy import FsdpGeneration
-
 
 @dataclass(frozen=True)
 class _BridgePrompt:
@@ -65,7 +63,6 @@ class _WorkloadState:
     lora_alpha: int
     target_modules: Any
     target_parameters: tuple[str, ...] | None
-    fsdp_generation: FsdpGeneration
     exclude_modules: str | None
     warmstart_adapter: str | None
 

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/rl/launch/checkpoints.py

@@ -16,7 +16,6 @@ import flash.engine.worker.verl.checkpoints as verl_checkpoints
 from flash._internal.fileio import reject_duplicate_keys
 from flash.adapters.artifacts import MAX_ATTEMPT_ID, has_loadable_adapter_weights
 from flash.engine.support.verl_checkpoint import FsdpCheckpointInspection
-from flash.engine.support.verl_policy import FsdpGeneration
 from flash.engine.worker.io.heartbeat import join_while_draining
 from flash.engine.worker.train.core.lifecycle.ledger import CheckpointLedger
 from flash.engine.worker.train.entry.backend_common import (
@@ -516,7 +515,6 @@ def _restore_verl_resume(
     local_dir: str,
     *,
     world_size: int,
-    expected_fsdp_generation: FsdpGeneration,
     required_steps: tuple[int, ...] = (),
 ) -> int:
     """stage this run's current-contract native checkpoint and its internal adapter store."""
@@ -551,7 +549,6 @@ def _restore_verl_resume(
 
     resume, inspection = verl_checkpoints.select_verl_resume_checkpoint(
         world_size=world_size,
-        expected_fsdp_generation=expected_fsdp_generation,
         accept=accept,
     )
     if not resume:
@@ -564,7 +561,6 @@ def _restore_verl_resume(
             local_dir,
             job_label="GRPO",
             world_size=world_size,
-            expected_fsdp_generation=expected_fsdp_generation,
             inspection=inspection,
         )
     referenced = accepted.get(resume)
@@ -579,7 +575,6 @@ def _restore_verl_resume(
         local_dir,
         job_label="GRPO",
         world_size=world_size,
-        expected_fsdp_generation=expected_fsdp_generation,
         inspection=inspection,
     )
 

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/rl/launch/inputs.py

@@ -36,7 +36,6 @@ from flash.engine.plan.steps import (
     rl_data_parallel_cards,
     validate_save_steps,
 )
-from flash.engine.support.verl_policy import _resolve_fsdp_generation
 from flash.engine.worker.io.heartbeat import liveness_heartbeat
 from flash.engine.worker.runtime.rng import backend_seed, seed_training_rngs
 from flash.engine.worker.train.core.child.glue import validate_glue_template
@@ -398,7 +397,6 @@ def _assemble_grpo_inputs(
     schedule,
     prompt_opened_thinking,
     warmstart_adapter,
-    fsdp_generation,
     pinned,
 ):
     return {
@@ -456,7 +454,6 @@ def _assemble_grpo_inputs(
         "lora_rank": options["lora_rank"],
         "lora_alpha": options["lora_alpha"],
         "warmstart_adapter": warmstart_adapter,
-        "fsdp_generation": fsdp_generation,
         "epochs": schedule["epochs"],
         "verl_total_epochs": schedule["verl_total_epochs"],
         "steps": schedule["steps"],
@@ -600,6 +597,5 @@ def _resolve_grpo_inputs():
         schedule=schedule,
         prompt_opened_thinking=prompt_opened_thinking,
         warmstart_adapter=warmstart_adapter,
-        fsdp_generation=_resolve_fsdp_generation("grpo", targeting.target_parameters),
         pinned=pinned,
     )

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/rl/launch/verl_config.py

@@ -303,7 +303,7 @@ def _actor_overrides(cfg: dict) -> list[str]:
         f"actor_rollout_ref.actor.fsdp_config.reshard_after_forward={_hydra_val(bool(cfg.get('reshard_after_forward', True)))}",
         # grpo requires fsdp2 for dense and fused-expert actors. reshard_after_forward above still
         # independently selects zero-2 or zero-3, so the allocator's memory decision is unchanged.
-        *actor_fsdp_strategy_overrides(cfg["fsdp_generation"]),
+        *actor_fsdp_strategy_overrides(),
         # store the frozen base in bf16, not verl's fp32 yaml default. shared with the opd driver.
         *trainer_dtype_overrides(),
     ]
@@ -639,7 +639,6 @@ def _build_verl_training_cfg(
         # the catalog id, never model_path: fused routed-expert parameters are part of the same
         # resolved target surface and must not be derived from the local snapshot path.
         "target_parameters": targeting.target_parameters,
-        "fsdp_generation": inp["fsdp_generation"],
         "multimodal": bool(inp.get("multimodal")),
         "lr": inp["lr"],
         "group_size": inp["group_size"],

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/sft/child/plugin.py

@@ -2,8 +2,6 @@
 
 from __future__ import annotations
 
-from importlib import import_module
-
 if __name__ == "flash_sft_plugin":
     import flash_verl_runtime as runtime
 else:
@@ -100,46 +98,6 @@ def _install_reentrant_checkpointing(*, multimodal: bool) -> None:
     FSDPEngine._build_module = build_reentrant_module
 
 
-def _install_loraplus(ratio: float, ready_marker: str) -> None:
-    if ratio <= 1:
-        return
-    from peft.optimizers import create_loraplus_optimizer
-    from verl.workers.engine.fsdp.transformer_impl import FSDPEngine
-
-    def build_loraplus_optimizer(self, module):
-        config = self.optimizer_config
-        optimizer_cls = getattr(import_module(config.optimizer_impl), config.optimizer)
-        optimizer_kwargs = {"lr": config.lr, "weight_decay": config.weight_decay}
-        if "adam" in config.optimizer.lower() or "ademamix" in config.optimizer.lower():
-            optimizer_kwargs["betas"] = config.betas
-        if config.override_optimizer_config:
-            optimizer_kwargs.update(config.override_optimizer_config)
-        try:
-            optimizer = create_loraplus_optimizer(
-                model=module,
-                optimizer_cls=optimizer_cls,
-                optimizer_kwargs=optimizer_kwargs,
-                loraplus_lr_ratio=ratio,
-                loraplus_weight_decay=config.weight_decay,
-            )
-        except TypeError:
-            optimizer = create_loraplus_optimizer(
-                model=module,
-                optimizer_cls=optimizer_cls,
-                lr=config.lr,
-                loraplus_lr_ratio=ratio,
-                loraplus_weight_decay=config.weight_decay,
-                **{key: value for key, value in optimizer_kwargs.items() if key != "lr"},
-            )
-        print(
-            f"{ready_marker} ratio={ratio:g} optimizer={optimizer_cls.__name__}",
-            flush=True,
-        )
-        return optimizer
-
-    FSDPEngine._build_optimizer = build_loraplus_optimizer
-
-
 def _install_core(config: dict) -> None:
     seed = int(config["seed"])
     _install_exact_dataloaders()
@@ -150,10 +108,6 @@ def _install_core(config: dict) -> None:
         runtime.install_checkpoint_handler_filter(save_at_steps, int(config["total_steps"]))
     if config.get("reentrant_gradient_checkpointing"):
         _install_reentrant_checkpointing(multimodal=bool(config.get("multimodal")))
-    _install_loraplus(
-        float(config.get("loraplus_ratio", 1.0)),
-        str(config["loraplus_ready_marker"]),
-    )
 
 
 def install() -> None:

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/train/sft/setup/config.py

@@ -2,18 +2,14 @@
 
 `build_sft_overrides` turns a resolved config dict into the hydra argv verl's sft trainer takes, and
 the `_render_*`/`render_*` helpers emit the child-side python that goes with it: the dataset module
-that reads flash's parquet, and the sitecustomize that patches verl in-process (tf32, gdn varlen,
-exact dataloader order, LoRA+). It is all string and dict construction -- no torch, no subprocess.
+that reads flash's parquet. Child runtime patches are installed by the external SFT plugin.
+It is all string and dict construction -- no torch, no subprocess.
 
 Split out of `flash.engine.worker.train.entry.sft_train` to keep that module under the file-size limit.
 """
 
 from __future__ import annotations
 
-# printed by the LoRA+ shim once it has installed itself, and matched on the child's stdout by the
-# parent: the shim runs inside verl, so this line is the only proof the optimizer grouping took.
-_LORAPLUS_READY_MARKER = "FLASH_LORAPLUS_READY"
-
 # every key `build_sft_overrides` requires its input to carry. an override the child never receives
 # is a silently-defaulted hyperparameter, so a missing key raises rather than falling back.
 _REQUIRED_OVERRIDE_KEYS = (
@@ -28,7 +24,6 @@ _REQUIRED_OVERRIDE_KEYS = (
     "lora_rank",
     "lora_alpha",
     "target_modules",
-    "fsdp_generation",
     "ulysses_sp_size",
     "lr",
     "warmup_ratio",
@@ -71,8 +66,7 @@ def _hydra_val(value) -> str:
 
 
 # verl imports this optimizer by name in the torch-bearing worker process.
-# use fp32 adamw: bitsandbytes cannot update fsdp2 dtensors. fsdp1 is not an alternative because flattened
-# names make peft's lora+ grouping treat every parameter as one-dimensional and silently mis-group it.
+# use fp32 adamw: bitsandbytes cannot update fsdp2 dtensors.
 _VERL_OPTIMIZER_IMPL = "torch.optim"
 _VERL_OPTIMIZER_NAME = "AdamW"
 
@@ -136,7 +130,7 @@ def build_sft_overrides(cfg: dict) -> list[str]:
         f"model.fused_kernel_options.impl_backend={cfg['fused_ce_backend']}",
         f"model.use_liger={_hydra_val(cfg.get('use_liger', False))}",
         f"model.enable_gradient_checkpointing={_hydra_val(cfg.get('gradient_checkpointing', True))}",
-        f"engine.strategy={'fsdp2' if cfg['fsdp_generation'] == 2 else 'fsdp'}",
+        "engine.strategy=fsdp2",
         "engine.model_dtype=bfloat16",
         f"engine.seed={_hydra_val(cfg['seed'])}",
         f"engine.ulysses_sequence_parallel_size={_hydra_val(cfg['ulysses_sp_size'])}",

{freesolo_flash_dev-1.2.133 → freesolo_flash_dev-1.2.134}/flash/engine/worker/verl/capabilities.py

@@ -17,7 +17,6 @@ import shutil
 import subprocess
 from collections.abc import Sequence
 
-from flash.engine.support.verl_policy import FsdpGeneration
 from flash.engine.worker.train.entry.backend_common import (
     CAUSAL_CONV1D_REQUIREMENT,
     FLA_REQUIREMENT,
@@ -865,10 +864,9 @@ def rollout_layered_summon_overrides(target_parameters: Sequence[str] | None) ->
     return ["actor_rollout_ref.rollout.layered_summon=true"]
 
 
-def actor_fsdp_strategy_overrides(fsdp_generation: FsdpGeneration) -> list[str]:
-    """render one explicit bare actor strategy override from the resolved private policy."""
-    strategy = "fsdp2" if fsdp_generation == 2 else "fsdp"
-    return [f"actor_rollout_ref.actor.strategy={strategy}"]
+def actor_fsdp_strategy_overrides() -> list[str]:
+    """use fsdp2 for every actor, independent of model and lora targets."""
+    return ["actor_rollout_ref.actor.strategy=fsdp2"]
 
 
 def rollout_mm_processor_cache_overrides() -> list[str]: