Merge remote-tracking branch 'upstream/main' into fp8-rl-mxfp8-weight-sync

# Conflicts:
#	skyrl/backends/skyrl_train/inference_servers/new_inference_worker_wrap.py

Author: YJHMITWEB

.claude/docs/ci.md

@@ -12,7 +12,7 @@
 
 - Compute config: `l4_ci` (referenced from `ci/anyscale_*.yaml`).
 - Cloud: `sky-anyscale-aws-us-east-1`.
-- Image: `novaskyai/skyrl-train-ray-2.56.0-py3.12-cu12.8` (varies per workflow).
+- Image: `novaskyai/skyrl-train-ray-2.57.0-py3.12-cu13.0` (varies per workflow).
 - Logs: visit the Anyscale job page linked from the GitHub Actions step output. Stderr from Ray workers shows up under the head node logs, not the entrypoint logs.
 
 ## Adding a New Test to CI

.claude/docs/tinker.md

@@ -5,6 +5,7 @@ SkyRL's implementation of the [Tinker API](https://tinker-docs.thinkingmachines.
 ## Code Layout
 
 - **`skyrl/tinker/api.py`** -- FastAPI HTTP server. Receives Tinker SDK requests, writes them to SQLite/Postgres, returns future IDs.
+- **`skyrl/tinker/proto_serialization.py`** -- Proto wire-format conversion. SDK >= 0.25.0 sends `forward_backward` request bodies as protobuf (with forward-only passes routed to `/forward_backward` via a `forward_only` flag instead of `/forward`) and requires proto-serialized `SampleResponse`/`ForwardBackwardOutput` results from `retrieve_future`. Mirrors the SDK's `tinker/proto/{request,response}_conv.py`; round-trip tested in `tests/tinker/test_proto_serialization.py`.
 - **`skyrl/tinker/engine.py`** -- Background subprocess (`TinkerEngine`). Polls DB, batches compatible requests, dispatches to backend.
 - **`skyrl/tinker/types.py`** -- Internal Pydantic models (distinct from API request/response models in `api.py`). `LOSS_TYPES` dict defines valid loss functions.
 - **`skyrl/tinker/config.py`** -- `EngineConfig` Pydantic model. `add_model()` auto-generates argparse flags from Pydantic fields.

.github/workflows/cpu_skyrl.yaml

@@ -37,8 +37,12 @@ jobs:
     steps:
     - uses: actions/checkout@v4
     - name: Install uv
-      run: |
-        curl -LsSf https://astral.sh/uv/install.sh | sh
+      # Use the action rather than the curl installer: on GitHub-hosted runners it
+      # also restores/saves ~/.cache/uv between runs, which matters here because
+      # this job resolves several distinct extra sets (tinker/jax/dev).
+      uses: astral-sh/setup-uv@v6
+      with:
+        enable-cache: true
     - name: Run lint
       run: |
         # Pin to match the ruff rev in .pre-commit-config.yaml. Unpinned `uvx ruff`

.github/workflows/cpu_skyrl_train.yaml

@@ -42,7 +42,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:
@@ -64,7 +63,6 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:
@@ -88,7 +86,6 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_ci_h100.yaml

@@ -52,7 +52,6 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_e2e_ci.yaml

@@ -26,7 +26,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_e2e_ci_fully_async.yaml

@@ -26,7 +26,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_e2e_ci_megatron.yaml

@@ -26,7 +26,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_e2e_ci_sft.yaml

@@ -26,7 +26,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with:

.github/workflows/gpu_e2e_ci_tinker.yaml

@@ -26,7 +26,6 @@ jobs:
         with:
           # Semantic version range syntax or exact version of a Python version
           python-version: '3.12'
-          cache: 'pip'
       - name: Install the latest version of uv
         uses: astral-sh/setup-uv@v6
         with: