feat(intern-decision-serve): Intern-Decision-4B behind semif-serve's HTTP surface

Contract, service and tests (fake engine, no GPU). Scores through the checkpoint's own inference.py (DecisionEngine.predict, sha256-pinned); maps semif decisions onto Jev choice questions, packs /decide/shared into calls of at most 16, runs orderings in waves, and keeps semif's error mapping, admission, body limit and hard VRAM cap. Deltas from semif-serve are listed in the contract.
2026-09-30 09:04:39 -07:00
parent bb806e3596
commit 5bbf0aaeba
18 changed files with 3196 additions and 0 deletions
@@ -0,0 +1,52 @@
+[project]
+name = "intern-decision-serve"
+version = "0.1.0"
+description = "Intern-Decision-4B (its own inference.py) behind semif-serve's HTTP surface"
+requires-python = ">=3.12"
+dependencies = [
+  "fastapi==0.118.0",
+  "uvicorn==0.37.0",
+]
+
+[project.optional-dependencies]
+# The real engine: the 2026-09-30 bench stack (semif-serve 0.1.4's torch/transformers plus what
+# the checkpoint's own inference.py imports: PIL and the Qwen3.5 processor, which needs torchvision).
+model = [
+  "torch==2.10.0",
+  "torchvision==0.25.0",
+  "transformers==5.17.0",
+  "pillow==12.3.0",
+]
+# Qwen3.5's fast kernels, as in the bench image (without them transformers runs its slower
+# reference PyTorch paths, and the bench numbers were measured with them).
+fast = [
+  "flash-linear-attention==0.5.2",
+  "causal-conv1d @ https://github.com/Dao-AILab/causal-conv1d/releases/download/v1.7.0/causal_conv1d-1.7.0+cu12torch2.10cxx11abiTRUE-cp312-cp312-linux_x86_64.whl ; sys_platform == 'linux' and platform_machine == 'x86_64'",
+]
+
+[dependency-groups]
+dev = ["pytest==8.4.2", "httpx==0.28.1"]
+
+[build-system]
+requires = ["setuptools>=68"]
+build-backend = "setuptools.build_meta"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.pytest.ini_options]
+testpaths = ["tests"]
+
+[[tool.uv.index]]
+name = "pytorch-cu128"
+url = "https://download.pytorch.org/whl/cu128"
+explicit = true
+
+[tool.uv.sources]
+torch = { index = "pytorch-cu128" }
+torchvision = { index = "pytorch-cu128" }
+
+[tool.uv]
+# Hold the transitive pins to the 2026-09-30 bench image (semif-serve:0.1.4's lock), so the
+# service runs the stack its acceptance numbers are compared against.
+constraint-dependencies = ["numpy==2.2.6", "huggingface-hub==1.31.0", "regex==2026.9.10", "tokenizers==0.23.2", "safetensors==0.8.0"]