[staging] PR-5351 cross-OS validation (doc extraction + 3 fixes)

.github/workflows/pr5351-cpu-inference-macos.yml

@@ -0,0 +1,52 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 CPU-inference cross-OS lane: macOS (Apple Silicon).
+# Same as the Ubuntu lane but on macos-14. llama-cpp-python builds
+# with Metal autodetect disabled to stay on the CPU code path so the
+# result mirrors a non-GPU Mac.
+
+name: PR-5351 CPU inference macOS
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/core/chat/**'
+      - 'tests/studio/test_cpu_inference_on_extracted_document.py'
+      - '.github/workflows/pr5351-cpu-inference-macos.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-cpu-inference-macos-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  cpu-inference:
+    runs-on: macos-14
+    timeout-minutes: 40
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend + llama-cpp-python (CPU build)
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install \
+            python-multipart aiofiles sqlalchemy cryptography \
+            pyyaml jinja2 mammoth pymupdf pymupdf4llm pytest pytest-asyncio \
+            pytest-timeout huggingface_hub requests numpy
+          # Disable Metal so the lane stays CPU-only; mirrors a no-GPU Mac.
+          CMAKE_ARGS="-DGGML_METAL=OFF -DGGML_ACCELERATE=OFF -DGGML_NATIVE=OFF" \
+            pip install --upgrade --quiet llama-cpp-python
+
+      - name: CPU inference on extracted document
+        env:
+          PR5351_LLAMA_THREADS: '3'
+        run: |
+          python -m pytest -q tests/studio/test_cpu_inference_on_extracted_document.py -s --tb=short

.github/workflows/pr5351-cpu-inference-ubuntu.yml

@@ -0,0 +1,53 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 CPU-inference cross-OS lane: Ubuntu.
+# Builds llama-cpp-python from source for CPU, downloads a 0.5B GGUF
+# from HF, extracts a synthetic PDF via the PR's document extractor,
+# and asserts the model answers a ground-truth question. Proves
+# end-to-end document-attach -> extract -> inference works on a CPU
+# runner with no GPU.
+
+name: PR-5351 CPU inference Ubuntu
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/core/chat/**'
+      - 'tests/studio/test_cpu_inference_on_extracted_document.py'
+      - '.github/workflows/pr5351-cpu-inference-ubuntu.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-cpu-inference-ubuntu-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  cpu-inference:
+    runs-on: ubuntu-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend + llama-cpp-python (CPU build)
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install \
+            python-multipart aiofiles sqlalchemy cryptography \
+            pyyaml jinja2 mammoth pymupdf pymupdf4llm pytest pytest-asyncio \
+            pytest-timeout huggingface_hub requests numpy
+          # CPU wheel ships pre-built on Linux; falls back to source if needed.
+          CMAKE_ARGS="-DGGML_NATIVE=OFF" pip install --upgrade --quiet llama-cpp-python
+
+      - name: CPU inference on extracted document
+        env:
+          PR5351_LLAMA_THREADS: '4'
+        run: |
+          python -m pytest -q tests/studio/test_cpu_inference_on_extracted_document.py -s --tb=short

.github/workflows/pr5351-cpu-inference-windows.yml

@@ -0,0 +1,49 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 CPU-inference cross-OS lane: Windows.
+# llama-cpp-python wheels exist for Windows; if pip falls back to
+# source, MSVC is preinstalled on windows-latest. CPU-only.
+
+name: PR-5351 CPU inference Windows
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/core/chat/**'
+      - 'tests/studio/test_cpu_inference_on_extracted_document.py'
+      - '.github/workflows/pr5351-cpu-inference-windows.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-cpu-inference-windows-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  cpu-inference:
+    runs-on: windows-latest
+    timeout-minutes: 40
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend + llama-cpp-python (CPU build)
+        shell: pwsh
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install python-multipart aiofiles sqlalchemy cryptography pyyaml jinja2 mammoth pymupdf pymupdf4llm pytest pytest-asyncio pytest-timeout huggingface_hub requests numpy
+          $env:CMAKE_ARGS = "-DGGML_NATIVE=OFF"
+          pip install --upgrade --quiet llama-cpp-python
+
+      - name: CPU inference on extracted document
+        shell: pwsh
+        env:
+          PR5351_LLAMA_THREADS: '4'
+        run: |
+          python -m pytest -q tests/studio/test_cpu_inference_on_extracted_document.py -s --tb=short

.github/workflows/pr5351-macos.yml

@@ -0,0 +1,60 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 cross-OS validation: macOS lane.
+# macos-14 (arm64). Validates the multiprocessing `spawn` path that
+# differs from Linux's default `fork`, the MLX detection branch in
+# core/chat/vlm_capability.py, and Safari/WebKit-relevant filesystem
+# behaviour. CPU-only; CUDA spoof auto-engages via tests/conftest.py.
+
+name: PR-5351 macOS
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/**'
+      - 'tests/studio/**'
+      - 'tests/conftest.py'
+      - '.github/workflows/pr5351-macos.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-macos-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  pytest:
+    runs-on: macos-14
+    timeout-minutes: 25
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend test dependencies (CPU only)
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install \
+            python-multipart aiofiles sqlalchemy cryptography \
+            pyyaml jinja2 mammoth unpdf requests \
+            'numpy<3' pytest pytest-asyncio httpx
+          pip install --index-url https://download.pytorch.org/whl/cpu 'torch>=2.4,<2.11'
+          pip install 'transformers>=4.51,<5.5'
+
+      - name: PR-5351 document tests (macOS spawn semantics)
+        working-directory: studio/backend
+        env:
+          # macOS's default start method is spawn; exercise the same
+          # config users see in production.
+          UNSLOTH_STUDIO_EXTRACT_CONCURRENCY: '2'
+        run: |
+          python -m pytest -q tests/test_chat_document_extraction.py tests/test_chat_document_routes.py tests/test_inference_worker.py tests/test_vision_cache.py tests/test_anthropic_messages.py tests/test_openai_tool_passthrough.py tests/test_models_get_model_config_case_resolution.py --tb=short
+
+      - name: PR-5351 regression tests + cancel timing
+        run: |
+          python -m pytest -q tests/studio/test_extractor_semaphore_leak.py tests/studio/test_html_independent_of_inference.py tests/studio/test_gguf_singleton_shared.py tests/studio/test_stream_cancel_registration_timing.py --tb=short

.github/workflows/pr5351-ubuntu.yml

@@ -0,0 +1,57 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 cross-OS validation: Ubuntu lane.
+# Runs the document-extraction tests, the cancellation-timing structural
+# test, and the three regression tests added in the fix commit against
+# Python 3.11 on ubuntu-latest. CPU-only; the existing tests/conftest.py
+# auto-installs the CUDA spoof so unsloth/unsloth_zoo device probes
+# return "cuda".
+
+name: PR-5351 Ubuntu
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/**'
+      - 'tests/studio/**'
+      - 'tests/conftest.py'
+      - '.github/workflows/pr5351-ubuntu.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-ubuntu-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  pytest:
+    runs-on: ubuntu-latest
+    timeout-minutes: 20
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend test dependencies (CPU only)
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install \
+            python-multipart aiofiles sqlalchemy cryptography \
+            pyyaml jinja2 mammoth unpdf requests \
+            'numpy<3' pytest pytest-asyncio httpx
+          pip install --index-url https://download.pytorch.org/whl/cpu 'torch>=2.4,<2.11'
+          pip install 'transformers>=4.51,<5.5'
+
+      - name: PR-5351 document tests
+        working-directory: studio/backend
+        run: |
+          python -m pytest -q tests/test_chat_document_extraction.py tests/test_chat_document_routes.py tests/test_inference_worker.py tests/test_vision_cache.py tests/test_anthropic_messages.py tests/test_openai_tool_passthrough.py tests/test_models_get_model_config_case_resolution.py --tb=short
+
+      - name: PR-5351 regression tests + cancel timing
+        run: |
+          python -m pytest -q tests/studio/test_extractor_semaphore_leak.py tests/studio/test_html_independent_of_inference.py tests/studio/test_gguf_singleton_shared.py tests/studio/test_stream_cancel_registration_timing.py --tb=short

.github/workflows/pr5351-windows.yml

@@ -0,0 +1,59 @@
+# SPDX-License-Identifier: AGPL-3.0-only
+# Copyright 2026-present the Unsloth AI Inc. team. All rights reserved.
+#
+# PR-5351 cross-OS validation: Windows lane.
+# windows-latest. Validates the multiprocessing `spawn` path
+# (mandatory on Windows), path normalisation, and EAGAIN-style
+# Process construction failures under load (the exact bug class the
+# semaphore-leak fix protects against). CPU-only; CUDA spoof
+# auto-engages via tests/conftest.py.
+
+name: PR-5351 Windows
+
+on:
+  push:
+    branches: [pr-5351-cross-os-validation]
+    paths:
+      - 'studio/backend/**'
+      - 'tests/studio/**'
+      - 'tests/conftest.py'
+      - '.github/workflows/pr5351-windows.yml'
+  workflow_dispatch:
+
+concurrency:
+  group: pr5351-windows-${{ github.ref }}
+  cancel-in-progress: true
+
+jobs:
+  pytest:
+    runs-on: windows-latest
+    timeout-minutes: 30
+    steps:
+      - uses: actions/checkout@v4
+
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11'
+          cache: 'pip'
+
+      - name: Install backend test dependencies (CPU only)
+        shell: pwsh
+        run: |
+          python -m pip install --upgrade pip
+          pip install -r studio/backend/requirements/studio.txt
+          pip install python-multipart aiofiles sqlalchemy cryptography pyyaml jinja2 mammoth unpdf requests "numpy<3" pytest pytest-asyncio httpx
+          pip install --index-url https://download.pytorch.org/whl/cpu "torch>=2.4,<2.11"
+          pip install "transformers>=4.51,<5.5"
+
+      - name: PR-5351 document tests (Windows spawn semantics)
+        working-directory: studio/backend
+        shell: pwsh
+        env:
+          UNSLOTH_STUDIO_EXTRACT_CONCURRENCY: '2'
+        run: |
+          python -m pytest -q tests/test_chat_document_extraction.py tests/test_chat_document_routes.py tests/test_inference_worker.py tests/test_vision_cache.py tests/test_anthropic_messages.py tests/test_openai_tool_passthrough.py tests/test_models_get_model_config_case_resolution.py --tb=short
+
+      - name: PR-5351 regression tests + cancel timing
+        shell: pwsh
+        run: |
+          python -m pytest -q tests/studio/test_extractor_semaphore_leak.py tests/studio/test_html_independent_of_inference.py tests/studio/test_gguf_singleton_shared.py tests/studio/test_stream_cancel_registration_timing.py --tb=short