fix(ci): install test dependencies before training pins (#3580)

.github/workflows/_run-ci.yml

@@ -166,6 +166,7 @@ jobs:
             v=$(python -m pip show "$pkg" 2>/dev/null | awk '/^Version:/{print $2}' || true)
             if [ -n "$v" ]; then pinned="$pinned $pkg==$v"; fi
           done
+          python -m pip install -r "$GITHUB_WORKSPACE/examples/multi_lora/requirements.txt" --break-system-packages
           python -m pip install -r "$GITHUB_WORKSPACE/requirements.txt" --break-system-packages
           for spec in $pinned; do
             pkg=${spec%%==*}

docs/developer/ci/00-stage.md

@@ -72,7 +72,7 @@ A test normally runs at its home stage. A [dispatch label](/developer/ci/01-labe
 
 Without a dispatch label nothing leaves its home stage, so scheduled and called runs are unaffected. Whatever stage actually executes a run is also the stage its performance baseline is keyed on, keeping the generations' numbers apart.
 
-**Dependency boundary.** CUDA stages start from dependencies baked into `radixark/miles`, reconcile Miles runtime dependencies from `requirements.txt`, update the SGLang and Megatron-LM checkouts to the selected refs, and expose all three source trees through `PYTHONPATH`; they do not rebuild or install the Miles, SGLang, or Megatron-LM source trees after the container starts. The hosted CPU stages install dependencies from `requirements.txt` and the fully pinned `tests/ci/requirements-ci-cpu.txt`, then expose the Miles, SGLang, and Megatron-LM source trees through `PYTHONPATH` without editable installs or inline package lists. The ROCm stage instead uses the SGLang and Megatron-LM versions baked into `rocm/sgl-dev`, unless the run names a ref for either.
+**Dependency boundary.** CUDA stages start from dependencies baked into `radixark/miles`, install CUDA test dependencies from `examples/multi_lora/requirements.txt`, override them with Miles runtime dependencies from `requirements.txt`, restore the image’s cuDNN pins, update the SGLang and Megatron-LM checkouts to the selected refs, and expose all three source trees through `PYTHONPATH`; they do not rebuild or install the Miles, SGLang, or Megatron-LM source trees after the container starts. The hosted CPU stages install dependencies from `requirements.txt` and the fully pinned `tests/ci/requirements-ci-cpu.txt`, then expose the Miles, SGLang, and Megatron-LM source trees through `PYTHONPATH` without editable installs or inline package lists. The ROCm stage instead uses the SGLang and Megatron-LM versions baked into `rocm/sgl-dev`, unless the run names a ref for either.
 
 CUDA and CPU dependency refs resolve in this order: explicit dispatch input or PR-body directive, committed `release-lock.json`, then the moving `sglang-miles` / `miles-main` branch heads. A called release run therefore checks out its requested Miles `ref` and consumes the lockfile on that ref unless an explicit override exists. ROCm checks out the requested Miles ref but keeps the dependencies baked into its image unless the run names a ref for one.
 

examples/multi_lora/requirements.txt

@@ -0,0 +1,2 @@
+tinker==0.26.2
+tinker_cookbook[math-rl] @ git+https://github.com/thinking-machines-lab/tinker-cookbook@1f962eda3a2c

examples/multi_lora/run_client_recipes.py

@@ -3,9 +3,9 @@
 The cookbook's sl_loop (SFT) and rl_loop (GRPO) are the executable definition
 of the Tinker wire contract; passing them is the gateway's acceptance bar.
 
-Requires, next to the pinned SDK (tinker==0.26.2):
+Install the client dependencies:
 
-    pip install git+https://github.com/thinking-machines-lab/tinker-cookbook@1f962eda3a2c
+    pip install -r examples/multi_lora/requirements.txt
 
 ``--base-model`` must be both the name this gateway serves (--tinker-base-model)
 and a HuggingFace name the cookbook can resolve a tokenizer and renderer for.

tests/e2e/lora/test_tinker_cookbook_gateway.py

@@ -12,13 +12,6 @@ register_cuda_ci(
     hardware=["hopper"],
 )
 
-COOKBOOK_PIN = "git+https://github.com/thinking-machines-lab/tinker-cookbook@1f962eda3a2c"
-
-
-def prepare():
-    prepare_gateway()
-    U.exec_command_cpu(f"pip install tinker==0.26.2 {COOKBOOK_PIN}")
-
 
 def execute():
     with running_gateway() as base_url:
@@ -29,5 +22,5 @@ def execute():
 
 
 if __name__ == "__main__":
-    prepare()
+    prepare_gateway()
     execute()

tests/e2e/lora/tinker_gateway.py

@@ -18,7 +18,6 @@ SERVE_TIMEOUT_S = 1200
 def prepare_gateway():
     U.exec_command_cpu("mkdir -p /root/models")
     U.exec_command_cpu(f"hf download {BASE_MODEL} --local-dir /root/models/{MODEL_NAME}")
-    U.exec_command_cpu("pip install tinker==0.26.2")
 
 
 def _wait_for_gateway(server: subprocess.Popen) -> None: