freesolo-flash-dev 1.4.7__tar.gz → 1.4.9__tar.gz

freesolo_flash_dev-1.4.9/.github/scripts/ci_sharding.py

@@ -0,0 +1,59 @@
+"""partition collected tests by whole file without changing pytest discovery."""
+
+from __future__ import annotations
+
+from collections.abc import Iterable
+
+import pytest
+
+# slow-file weights come from offline junit timings; unmeasured files still participate.
+FILE_SECONDS = {
+    "tests/test_client.py": 23,
+    "tests/test_instance_bootstrap_coverage.py": 6,
+    "tests/test_repo_cleanup.py": 24,
+    "tests/test_runmgmt.py": 16,
+    "tests/test_server_api.py": 39,
+    "tests/test_server_envs_coverage.py": 15,
+    "tests/test_weight_cache.py": 7,
+    "tests/test_worker_init_heartbeat.py": 4,
+}
+
+
+def assign_files(paths: Iterable[str], count: int) -> dict[str, int]:
+    loads = [0.0] * count
+    assignments = {}
+    for path in sorted(paths, key=lambda path: (-FILE_SECONDS.get(path, 1.0), path)):
+        shard = min(range(count), key=lambda index: (loads[index], index))
+        assignments[path] = shard
+        loads[shard] += FILE_SECONDS.get(path, 1.0)
+    return assignments
+
+
+def pytest_addoption(parser):
+    group = parser.getgroup("ci sharding")
+    group.addoption("--ci-shard", type=int, default=None)
+    group.addoption("--ci-shard-count", type=int, default=2)
+
+
+def pytest_configure(config):
+    shard = config.getoption("--ci-shard")
+    count = config.getoption("--ci-shard-count")
+    if shard is None or count < 1 or not 0 <= shard < count:
+        raise pytest.UsageError("ci shard must satisfy 0 <= --ci-shard < --ci-shard-count")
+
+
+@pytest.hookimpl(wrapper=True)
+def pytest_collection_modifyitems(config, items):
+    # partition after marker selection so every worker assigns the same eligible file set.
+    result = yield
+    count = config.getoption("--ci-shard-count")
+    shard = config.getoption("--ci-shard")
+    paths = {item.path.relative_to(config.rootpath).as_posix() for item in items}
+    assignments = assign_files(paths, count)
+    kept, removed = [], []
+    for item in items:
+        path = item.path.relative_to(config.rootpath).as_posix()
+        (kept if assignments[path] == shard else removed).append(item)
+    config.hook.pytest_deselected(items=removed)
+    items[:] = kept
+    return result

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/.github/workflows/ci.yml

@@ -16,19 +16,10 @@ concurrency:
   cancel-in-progress: true
 
 jobs:
-  test:
-    name: lint + offline tests (py${{ matrix.python-version }})
+  lint:
+    name: lint
     runs-on: ubuntu-latest
-    # The suite is CPU-only and finishes well inside this; the cap is here so a wedged run
-    # releases its runner instead of holding it for the 6-hour default.
-    timeout-minutes: 30
-    strategy:
-      fail-fast: false
-      matrix:
-        # pyproject declares `requires-python = ">=3.11,<3.13"`, so both supported interpreters
-        # are tested. Testing only 3.11 let a 3.12-only failure reach a user the project claims
-        # to support.
-        python-version: ["3.11", "3.12"]
+    timeout-minutes: 10
     steps:
       - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
 
@@ -37,7 +28,7 @@ jobs:
         with:
           # pin the installed binary as well as the action implementation.
           version: "0.12.5"
-          python-version: ${{ matrix.python-version }}
+          python-version: "3.11"
 
       - name: Sync (server extra + dev)
         run: uv sync --extra server --extra serving --dev --locked
@@ -64,69 +55,68 @@ jobs:
       - name: Source layout check
         run: uv run python scripts/check_source_layout.py
 
-      # The job has always been NAMED offline (tests/conftest.py stubs the network boundaries in
-      # Python), but nothing enforced it, so a test that reached hf.co passed on a connected
-      # runner and would go red on a blip. Setting these makes the name true: an unstubbed call
-      # fails locally and deterministically instead of intermittently in CI.
-      #
-      # Sharding: this step is ~300s of the job's ~305s, and almost all of it is waiting rather
-      # than computing (the offline tests exercise retry/backoff/watchdog paths, so 125 sleep call
-      # sites sit idle holding a core). One process left the runner's second core idle for five
-      # minutes; `-n 2` cuts the step roughly in half.
-      #
-      # `loadfile` (not the default `load`, and not `loadscope`) is required, not a preference:
-      # tests in the same FILE share module-level import state -- several re-import
-      # `flash.engine.worker` to re-run its import scope, and others patch dotted targets under it.
-      # Splitting a file across workers breaks that sharing; `loadscope` splits at class/module
-      # granularity and fails tests/test_warmstart_cross_repo.py for exactly this reason. `loadfile`
-      # keeps every file whole on one worker, which is the boundary the suite already assumes.
-      #
-      # `-n 2` is pinned deliberately instead of `-n auto`. `auto` does NOT mean "the cores this
-      # process may use": xdist prefers psutil.cpu_count() when psutil is importable (it is, via a
-      # transitive dep) and that ignores CPU affinity and cgroup limits. So `auto` silently couples
-      # the worker count to a transitive dependency and to whatever tier the runner pool hands out.
-      # A standard GitHub runner is 2-core, which is what this suite was measured against; pinning
-      # keeps the sharding identical on every run instead of changing under us.
+      # The offline suite can only compare the pin against a hardcoded copy of itself, so a
+      # mis-transcribed sha passes every pin test while naming a commit that was never pushed --
+      # which then fails EVERY worker install at provisioning time. This is the one pin check that
+      # asks upstream instead of asking the repo about itself, so it needs the network.
       #
-      # Oversubscription is not free here: at 48 workers on 2 cores, three timing-sensitive tests in
-      # tests/test_server_envs_coverage.py fail, because they assert on a schema-validation timeout
-      # budget that CPU starvation blows through. Raise this only alongside the runner's core count.
+      # Offline environment variables are scoped to pytest steps; this check needs the network.
+      - name: verl pin exists upstream
+        run: uv run python scripts/check_verl_pin_exists.py
+
+  test:
+    name: offline tests (py${{ matrix.python-version }}, shard ${{ matrix.shard }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    strategy:
+      fail-fast: false
+      matrix:
+        python-version: ["3.11", "3.12"]
+        shard: [0, 1]
+    steps:
+      - uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@1c37ad07a6a961277cf70c0d37d6f313000f5884 # v10.3.0
+        with:
+          # pin the installed binary as well as the action implementation.
+          version: "0.12.5"
+          python-version: ${{ matrix.python-version }}
+
+      - name: Sync (server extra + dev)
+        run: uv sync --extra server --extra serving --dev --locked
+
+      # Keep each file whole and each shard single-process for shared module state.
       - name: Tests (CPU, offline)
         env:
           HF_HUB_OFFLINE: "1"
           TRANSFORMERS_OFFLINE: "1"
           HF_HUB_DISABLE_TELEMETRY: "1"
-        run: uv run pytest -q -n 2 --dist loadfile -m "not wallclock"
-
-      # The `wallclock` tests assert on real elapsed time with tens-of-milliseconds margins, so a
-      # second worker competing for the core fails them for reasons unrelated to the code. They are
-      # deselected above and run here alone, single-process: ~5s, which is why isolating them is
-      # cheaper than the scaled-up timing budgets that would be needed to survive contention.
-      - name: Tests (CPU, offline, wall-clock sensitive)
+          PYTHONPATH: .github/scripts
+        run: >-
+          uv run pytest -q -p ci_sharding --ci-shard ${{ matrix.shard }}
+          --ci-shard-count 2 -m "not wallclock" --durations=20
+
+      # Run timing-sensitive tests separately and exactly once per Python version.
+      # Shard zero also retains the real CPU tensor regressions after dependency installation.
+      - if: matrix.shard == 0
+        name: Tests (CPU, offline, wall-clock sensitive)
         env:
           HF_HUB_OFFLINE: "1"
           TRANSFORMERS_OFFLINE: "1"
           HF_HUB_DISABLE_TELEMETRY: "1"
         run: uv run pytest -q -m wallclock
 
-      # The offline suite can only compare the pin against a hardcoded copy of itself, so a
-      # mis-transcribed sha passes every pin test while naming a commit that was never pushed --
-      # which then fails EVERY worker install at provisioning time. This is the one pin check that
-      # asks upstream instead of asking the repo about itself, so it needs the network.
-      #
-      # Deliberately NOT covered by the offline env above: those vars are scoped to the pytest step,
-      # so this step still reaches github as it must.
-      - name: verl pin exists upstream
-        run: uv run python scripts/check_verl_pin_exists.py
-
       # The ordinary offline environment omits torch/PEFT, so its adapter tests skip.
       # Exercise real meta tensors and frozen references with the worker's versions.
-      - name: Install CPU adapter-test dependencies
+      - if: matrix.shard == 0
+        name: Install CPU adapter-test dependencies
         run: |
           uv pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu
           uv pip install peft==0.20.0 transformers==5.12.1
 
-      - name: Warm-start KL reference regression (real torch and PEFT)
+      - if: matrix.shard == 0
+        name: Warm-start KL reference regression (real torch and PEFT)
         env:
           HF_HUB_OFFLINE: "1"
           TRANSFORMERS_OFFLINE: "1"
@@ -135,32 +125,13 @@ jobs:
           .venv/bin/python -c "import torch, peft"
           .venv/bin/pytest -q tests/test_rl_train.py -k kl_ref_adapter
 
-      - name: Custom training gradient and session regressions (real CPU torch)
+      - if: matrix.shard == 0
+        name: Custom training gradient and session regressions (real CPU torch)
         env:
           HF_HUB_OFFLINE: "1"
           TRANSFORMERS_OFFLINE: "1"
         run: .venv/bin/pytest -q tests/test_custom_gradients.py tests/test_custom_session.py tests/test_custom_client.py tests/test_custom_managed.py
 
-  # `lint + offline tests` is the required status check, and the matrix above turned that one
-  # check into two per-version checks. This aggregate keeps a single stable name for branch
-  # protection to require, and fails if either matrix leg failed or was cancelled.
-  test-result:
-    name: lint + offline tests
-    runs-on: ubuntu-latest
-    needs: test
-    if: always()
-    timeout-minutes: 5
-    steps:
-      - name: Require every matrix leg to have succeeded
-        env:
-          RESULT: ${{ needs.test.result }}
-        run: |
-          if [ "$RESULT" != "success" ]; then
-            echo "::error::matrix job 'test' finished with result '$RESULT'."
-            exit 1
-          fi
-          echo "All matrix legs passed."
-
   typecheck:
     name: mypy (advisory)
     runs-on: ubuntu-latest
@@ -222,3 +193,23 @@ jobs:
             } >> "$GITHUB_STEP_SUMMARY"
             echo "::warning::mypy stopped early; the reported error count is not the real total."
           fi
+
+  test-result:
+    name: lint + offline tests
+    runs-on: ubuntu-latest
+    timeout-minutes: 5
+    needs: [lint, test]
+    if: always()
+    steps:
+      - name: Require every test group to have succeeded
+        env:
+          LINT: ${{ needs.lint.result }}
+          TEST: ${{ needs.test.result }}
+        run: |
+          for result in "$LINT" "$TEST"; do
+            if [ "$result" != "success" ]; then
+              echo "::error::required ci group finished with result '$result'."
+              exit 1
+            fi
+          done
+          echo "all required ci groups passed"

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/.github/workflows/publish-dev.yml

@@ -19,6 +19,7 @@ on:
       - "uv.lock"
       - "flash/**"
       - "scripts/build_dev_dist.py"
+      - "scripts/check_dev_release.py"
       - ".github/workflows/publish-dev.yml"
   workflow_dispatch:
 
@@ -82,24 +83,7 @@ jobs:
         id: decide
         env:
           VERSION: ${{ steps.meta.outputs.version }}
-        run: |
-          python3 - <<'PY' >> "$GITHUB_OUTPUT"
-          import os
-          import urllib.error
-          import urllib.request
-
-          version = os.environ["VERSION"]
-          url = f"https://pypi.org/pypi/freesolo-flash-dev/{version}/json"
-          published = False
-          try:
-              with urllib.request.urlopen(url, timeout=30) as response:
-                  published = response.status == 200
-          except urllib.error.HTTPError as error:
-              if error.code != 404:
-                  raise
-          # Publish only when the dev version isn't on PyPI yet (i.e. it was just bumped).
-          print(f"publish={'false' if published else 'true'}")
-          PY
+        run: python3 scripts/check_dev_release.py >> "$GITHUB_OUTPUT"
 
       - name: Already published (nothing to do)
         if: steps.decide.outputs.publish == 'false'

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/Dockerfile

@@ -18,11 +18,9 @@
 FROM python:3.12-slim@sha256:2c941e860699f878900b0edc2403613c234d4b32eda3cc9fa7036991a2a63c4a
 
 WORKDIR /app
-COPY . .
 RUN apt-get update \
     && apt-get install -y --no-install-recommends ca-certificates git curl \
     && rm -rf /var/lib/apt/lists/*
-RUN pip install --no-cache-dir ".[server]"
 
 # Optional Infisical CLI, for deployments that pull secrets from a vault at container start
 # instead of receiving them as environment variables. OFF by default: a plain `docker build .`
@@ -67,6 +65,10 @@ RUN if [ "${INSTALL_INFISICAL}" = "true" ]; then \
         infisical --version; \
     fi
 
+# Source changes reuse the operating-system and optional Infisical layers.
+COPY . .
+RUN pip install --no-cache-dir ".[server]"
+
 VOLUME /root/.flash
 EXPOSE 8080
 

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: freesolo-flash-dev
-Version: 1.4.7
+Version: 1.4.9
 Summary: Flash — managed LoRA post-training (SFT/GRPO) 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
@@ -144,11 +144,19 @@ path is self-hostable end to end — see **[SELF_HOSTING.md](SELF_HOSTING.md)**.
 
 ## Using the hosted service
 
-The active training catalog is `Qwen/Qwen3.5-9B`, `Qwen/Qwen3.8-27B`, and
-`Qwen/Qwen3.6-35B-A3B`. The hosted serving catalog routes all three models on B200.
-The retired dense Qwen3.6 27B identity is never translated into Qwen3.8. Customer-owned
-Modal supports all three active models; 27B and 35B-A3B are bound to the certified serving image
-digest. Customer-owned `flash serve` requires explicit `--provider modal`. Training and serving both use Modal.
+The training catalog includes Qwen3.5 0.8B, 4B, 9B, and **9B Base**, Qwen3.8 27B,
+and Qwen3.6 35B-A3B. `Qwen/Qwen3.5-9B` is Qwen's post-trained chat model and remains
+the default. Select `model = "Qwen/Qwen3.5-9B-Base"` to fine-tune the pretrained-only
+checkpoint with SFT or GRPO. It has its own weights, tokenizer revision, adapters,
+and Modal cache entry; it is never mapped to the post-trained model's FP8 weights.
+Its native chat template supports text, images, and tool calls, but the pretrained
+weights have not been tuned for instruction following.
+
+Hosted serving uses the [per-model GPU placements](flash/serving/app/README.md#hosted-gpu-configuration).
+The 9B Base checkpoint serves in BF16 on H100 with a 262,144-token context.
+Customer-owned `flash serve` requires explicit `--provider modal` and a live-qualified
+profile. The retired dense Qwen3.6 27B identity is never translated into Qwen3.8.
+Training and serving both use Modal.
 
 `flash login` is not interactive. Export `FREESOLO_API_KEY` first so the key does not appear in process listings:
 
@@ -348,6 +356,12 @@ and unions on top-level function arguments are unsupported. Top-level arguments
 native parameter text, where a string such as `"12"` and a number would be ambiguous;
 nested JSON preserves those types during generation, streaming, and history replay.
 
+Qualified Qwen tool-calling routes support thinking-enabled generation. Tool calls are
+parsed only after the first `</think>`; tool-like text inside reasoning stays in
+`reasoning_content`. A response that ends before that boundary cannot execute tools.
+Streaming preserves the same separation, and saved adapters retain their configured
+thinking mode. Tools still cannot be combined with logprobs or structured outputs.
+
 ### Workload estimates (SFT)
 
 `flash train`, `--dry-run`, and `--cost` read the pinned environment's packaged dataset

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/README.md

@@ -44,11 +44,19 @@ path is self-hostable end to end — see **[SELF_HOSTING.md](SELF_HOSTING.md)**.
 
 ## Using the hosted service
 
-The active training catalog is `Qwen/Qwen3.5-9B`, `Qwen/Qwen3.8-27B`, and
-`Qwen/Qwen3.6-35B-A3B`. The hosted serving catalog routes all three models on B200.
-The retired dense Qwen3.6 27B identity is never translated into Qwen3.8. Customer-owned
-Modal supports all three active models; 27B and 35B-A3B are bound to the certified serving image
-digest. Customer-owned `flash serve` requires explicit `--provider modal`. Training and serving both use Modal.
+The training catalog includes Qwen3.5 0.8B, 4B, 9B, and **9B Base**, Qwen3.8 27B,
+and Qwen3.6 35B-A3B. `Qwen/Qwen3.5-9B` is Qwen's post-trained chat model and remains
+the default. Select `model = "Qwen/Qwen3.5-9B-Base"` to fine-tune the pretrained-only
+checkpoint with SFT or GRPO. It has its own weights, tokenizer revision, adapters,
+and Modal cache entry; it is never mapped to the post-trained model's FP8 weights.
+Its native chat template supports text, images, and tool calls, but the pretrained
+weights have not been tuned for instruction following.
+
+Hosted serving uses the [per-model GPU placements](flash/serving/app/README.md#hosted-gpu-configuration).
+The 9B Base checkpoint serves in BF16 on H100 with a 262,144-token context.
+Customer-owned `flash serve` requires explicit `--provider modal` and a live-qualified
+profile. The retired dense Qwen3.6 27B identity is never translated into Qwen3.8.
+Training and serving both use Modal.
 
 `flash login` is not interactive. Export `FREESOLO_API_KEY` first so the key does not appear in process listings:
 
@@ -248,6 +256,12 @@ and unions on top-level function arguments are unsupported. Top-level arguments
 native parameter text, where a string such as `"12"` and a number would be ambiguous;
 nested JSON preserves those types during generation, streaming, and history replay.
 
+Qualified Qwen tool-calling routes support thinking-enabled generation. Tool calls are
+parsed only after the first `</think>`; tool-like text inside reasoning stays in
+`reasoning_content`. A response that ends before that boundary cannot execute tools.
+Streaming preserves the same separation, and saved adapters retain their configured
+thinking mode. Tools still cannot be combined with logprobs or structured outputs.
+
 ### Workload estimates (SFT)
 
 `flash train`, `--dry-run`, and `--cost` read the pinned environment's packaged dataset

freesolo_flash_dev-1.4.9/flash/content/processor.py

@@ -0,0 +1,17 @@
+"""Normalize chat-template metadata from a checkpoint's own processor and tokenizer."""
+
+from typing import Any
+
+
+def inherit_tokenizer_chat_template(processor: Any) -> None:
+    """Support checkpoints that publish their template only in tokenizer_config.json.
+
+    Transformers loads that template onto the tokenizer but not AutoProcessor.
+    Preserve explicit processor templates and never substitute another model's tokenizer.
+    Missing templates remain missing so the normal request validation can reject them.
+    """
+    if getattr(processor, "chat_template", None) is None:
+        tokenizer = getattr(processor, "tokenizer", None)
+        template = getattr(tokenizer, "chat_template", None)
+        if template is not None:
+            processor.chat_template = template

freesolo_flash_dev-1.4.9/flash/content/stop_tokens.py

@@ -0,0 +1,13 @@
+"""Chat-turn boundaries for pretrained checkpoints with document-level EOS metadata."""
+
+from typing import Any
+
+
+def chat_stop_token_ids(model: str, tokenizer: Any, *, is_chat: bool) -> list[int] | None:
+    """Stop Base chat at im_end without changing its tokenizer or raw completion EOS."""
+    if model != "Qwen/Qwen3.5-9B-Base" or not is_chat:
+        return None
+    token_id = tokenizer.convert_tokens_to_ids("<|im_end|>")
+    if type(token_id) is not int or token_id < 0 or token_id == tokenizer.unk_token_id:
+        raise ValueError("Qwen3.5 9B Base tokenizer is missing its chat end token")
+    return [token_id]

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/flash/core/catalog.py

@@ -16,6 +16,7 @@ _ON_POLICY_ALGORITHMS = frozenset({"grpo"})
 _IMAGE_TRAINING_MODELS = frozenset(
     {
         "Qwen/Qwen3.5-9B",
+        "Qwen/Qwen3.5-9B-Base",
         "Qwen/Qwen3.8-27B",
         "Qwen/Qwen3.6-35B-A3B",
     }
@@ -218,6 +219,7 @@ SERVING_MODEL_REPOS: dict[str, str] = {
     "Qwen/Qwen3.5-0.8B": "Qwen/Qwen3.5-0.8B",
     "Qwen/Qwen3.5-4B": "Qwen/Qwen3.5-4B",
     "Qwen/Qwen3.5-9B": "Freesolo-Co/Qwen3.5-9B-FP8",
+    "Qwen/Qwen3.5-9B-Base": "Qwen/Qwen3.5-9B-Base",
     "Qwen/Qwen3.8-27B": "Qwen/Qwen3.8-27B-FP8",
     "Qwen/Qwen3.6-35B-A3B": "Qwen/Qwen3.6-35B-A3B",
 }
@@ -532,6 +534,32 @@ MODELS: dict[str, ModelInfo] = {
 }
 
 
+# The pretrained checkpoint has the same architecture and native chat template as 9B,
+# but distinct weights, EOS metadata, and adapter identity. Never serve it via the
+# post-trained model's FP8 repository.
+MODELS["Qwen/Qwen3.5-9B-Base"] = replace(
+    MODELS["Qwen/Qwen3.5-9B"],
+    id="Qwen/Qwen3.5-9B-Base",
+    display_name="Qwen3.5 9B Base (pretrained)",
+    managed_revision="68c46c4b3498877f3ef123c856ecfde50c39f404",
+    serving=ServingCapacity(
+        gpu="H100",
+        serve_model_id="Qwen/Qwen3.5-9B-Base",
+        max_loras=16,
+        max_cpu_loras=16,
+        max_lora_rank=128,
+        max_model_len=262144,
+        max_num_seqs=8,
+        tensor_parallel_size=1,
+        gpu_memory_utilization=0.90,
+        image_limit=MAX_IMAGES_PER_CONVERSATION,
+    ),
+    notes="Pretrained-only Qwen3.5 9B; bf16 LoRA SFT/GRPO. Uses its own weights and "
+    "tokenizer, not the post-trained 9B. Native chat/thinking formatting is available for "
+    "fine-tuning; instruction following and reasoning are not post-trained. BF16 serving on H100.",
+)
+
+
 def list_models() -> list[ModelInfo]:
     return sorted(MODELS.values(), key=lambda m: (m.min_vram_gb, m.id))
 

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/flash/custom/backend.py

@@ -308,6 +308,7 @@ def load_backend(
     seed=42,
     checkpoint_dir,
     max_tokens=2048,
+    init_adapter=None,
 ):
     import torch
     from peft import LoraConfig, get_peft_model
@@ -342,17 +343,47 @@ def load_backend(
         torch_dtype=torch.bfloat16,
         attn_implementation="sdpa",
     ).to("cuda")
-    model = get_peft_model(
-        model,
-        LoraConfig(
-            r=lora_rank,
-            lora_alpha=lora_alpha or 2 * lora_rank,
-            lora_dropout=0,
-            target_modules="all-linear",
-            task_type="CAUSAL_LM",
-            revision=revision,
-            **targeting,
-        ),
-    )
+    if init_adapter:
+        model = _warm_start(model, init_adapter, lora_rank, lora_alpha)
+    else:
+        model = get_peft_model(
+            model,
+            LoraConfig(
+                r=lora_rank,
+                lora_alpha=lora_alpha or 2 * lora_rank,
+                lora_dropout=0,
+                target_modules="all-linear",
+                task_type="CAUSAL_LM",
+                revision=revision,
+                **targeting,
+            ),
+        )
     model.config._commit_hash = revision
     return TorchBackend(model, tokenizer, checkpoint_dir=checkpoint_dir, max_tokens=max_tokens)
+
+
+def _warm_start(model, adapter_dir, lora_rank, lora_alpha):
+    """Continue a verified source LoRA: its exact modules and weights, with fresh optimizer state.
+
+    The source config keeps its target modules; dropout is disabled like every custom adapter so
+    gradient checkpointing stays exact. Every source tensor must land on a trainable parameter:
+    peft only warns about unmatched keys, which would silently restart from zero.
+    """
+    from peft import LoraConfig, get_peft_model, load_peft_weights, set_peft_model_state_dict
+
+    config = LoraConfig.from_pretrained(adapter_dir)
+    if (config.r, config.lora_alpha) != (lora_rank, lora_alpha or 2 * lora_rank):
+        raise ValueError("warm-start adapter rank or alpha differs from the prepared run")
+    config.lora_dropout = 0
+    config.inference_mode = False
+    model = get_peft_model(model, config)
+    weights = load_peft_weights(adapter_dir, device="cuda")
+    result = set_peft_model_state_dict(model, weights)
+    trainable = [name for name, p in model.named_parameters() if p.requires_grad]
+    if result.unexpected_keys or len(weights) != len(trainable):
+        raise ValueError(
+            "warm-start adapter weights do not match the model's LoRA parameters "
+            f"({len(weights)} source tensors, {len(trainable)} trainable, "
+            f"{len(result.unexpected_keys)} unmatched)"
+        )
+    return model

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/flash/custom/worker.py

@@ -13,10 +13,34 @@ from flash.engine.worker.io import hf
 from flash.engine.worker.io.heartbeat import heartbeat, liveness_heartbeat
 
 
+def _custom_targeting(model_id):
+    """Custom runs adapt only text modules; a warm-start source must have the same modality."""
+    from flash.core.catalog import supports_image_training
+
+    if not supports_image_training(model_id):
+        return None
+    from flash.adapters.targets import resolve_lora_targeting
+
+    return resolve_lora_targeting(model_id, algorithm="custom", multimodal=False)
+
+
 def run(*, context):
     spec = context.config.spec
     heartbeat("custom_model_load", context=context)
     with liveness_heartbeat("custom_model_load", context=context):
+        init_adapter = None
+        if spec.train.init_from_adapter:
+            from flash.engine.worker.train.entry.sft_train import _warmstart_adapter_path
+
+            # downloads the prepared source and checks its rank, alpha, modules and base model.
+            init_adapter = _warmstart_adapter_path(
+                spec.model,
+                spec.model_revision,
+                spec.train.lora_rank,
+                spec.train.lora_alpha,
+                _custom_targeting(spec.model),
+                context=context,
+            )
         backend = load_backend(
             spec.model,
             revision=spec.model_revision,
@@ -25,6 +49,7 @@ def run(*, context):
             seed=spec.seed,
             checkpoint_dir="/tmp/flash-custom-checkpoints",
             max_tokens=spec.train.max_context_tokens,
+            init_adapter=init_adapter,
         )
     checkpoints = RunCheckpoints(context)
     backend.checkpoints = checkpoints

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/flash/engine/plan/vram.py

@@ -138,6 +138,7 @@ _SFT_CHUNKED_NLL_MODELS = frozenset(
         "Qwen/Qwen3.5-0.8B",
         "Qwen/Qwen3.5-4B",
         "Qwen/Qwen3.5-9B",
+        "Qwen/Qwen3.5-9B-Base",
         "Qwen/Qwen3.8-27B",
         "Qwen/Qwen3.6-35B-A3B",
     }

{freesolo_flash_dev-1.4.7 → freesolo_flash_dev-1.4.9}/flash/engine/profiling/sft_workload.py

@@ -12,6 +12,7 @@ from pathlib import Path
 from typing import Any, Protocol, TypedDict, cast
 
 from flash.content.multimodal import NormalizedImages
+from flash.content.processor import inherit_tokenizer_chat_template
 from flash.content.thinking import messages_for_chat_template
 from flash.engine.plan.recipe import RECIPE
 from flash.engine.plan.steps import resolve_update_horizon, sft_update_steps
@@ -190,11 +191,11 @@ def _default_processor_loader(model_id: str, revision: str):
 
     from flash.engine.support.huggingface import model_revision_kwargs
 
-    return AutoProcessor.from_pretrained(
-        model_id,
-        trust_remote_code=True,
-        **model_revision_kwargs(revision),
+    processor = AutoProcessor.from_pretrained(
+        model_id, trust_remote_code=True, **model_revision_kwargs(revision)
     )
+    inherit_tokenizer_chat_template(processor)
+    return processor
 
 
 def _packing_mode(