sync(slime #2014..#2125): 3-way merge [WIP — 45 conflict files need review]

.buildkite/gpu_suites.py

@@ -47,7 +47,6 @@
     "short": [
         ("test_qwen3.5_0.8B_gsm8k_async_short.py", 4, "", {}),
         ("test_qwen3.5_0.8B_gsm8k_short.py", 4, "", {}),
-        ("test_qwen2.5_0.5B_ppo_critic_only_short.py", 4, "", {}),
         ("test_qwen2.5_0.5B_fully_async_short.py", 4, "", {}),
     ],
     "vllm-config": [

.buildkite/pipeline.yml

@@ -78,6 +78,8 @@ steps:
           python tests/test_metric_report_dist.py
           python tests/test_loss_cp_invariance.py
           python tests/test_sample.py
+          python tests/test_cispo_loss.py
+          python tests/test_logprob_response_spans.py
           python tests/utils/test_hf_checkpoint_saver.py
         '
 
@@ -101,9 +103,10 @@ steps:
           pip install -q pytest numpy packaging pyyaml omegaconf tqdm httpx requests ray pybase64 pylatexenc sympy aiohttp pillow safetensors transformers cloudpickle
           pip install -q openai openai-agents anthropic
           pip install -q -e . --no-deps
-          python tests/test_agent_trajectory.py
-          python tests/test_agent_adapters.py
-          python tests/test_agent_sdk_adapters.py
+          python tests/test_agent/test_adapters.py
+          python tests/test_agent/test_harness.py
+          python tests/test_agent/test_trajectory_manager_branching.py
+          python tests/test_agent/test_agent_rollout_cpu.py
         '
 
   - label: ":pytest: utils tests"

.claude/skills/add-tests-and-ci/SKILL.md

@@ -40,15 +40,33 @@ if __name__ == "__main__":
 - `run-ci-changed` extracts a top-level `NUM_GPUS = <N>` constant from added/modified `tests/test_*.py` and `tests/plugin_contracts/test_*.py`; if missing, it defaults to 8 GPUs. Set `NUM_GPUS = 0` for CPU-only tests.
 - For GPU/e2e tests, follow the nearby file pattern (`prepare()`, `execute()`, `NUM_GPUS`, and any model/dataset constants).
 
-### Step 3: Run Local Validation
+### Step 3: Register Tests in GitHub CI
+
+Whenever adding, moving, or renaming a test file, update the GitHub workflow template before finishing:
+
+1. Add the test to the appropriate matrix in `.github/workflows/pr-test.yml.j2`.
+   - CPU-only pytest/unit tests usually belong in `cpu-unittest` with `num_gpus: 0`.
+   - GPU/e2e tests should be placed beside the nearest similar model/path test with the matching `num_gpus` and environment fields.
+2. Regenerate workflows:
+
+```bash
+python .github/workflows/generate_github_workflows.py
+```
+
+3. Include both `.github/workflows/pr-test.yml.j2` and the generated `.github/workflows/pr-test.yml` in the change set.
+
+Only skip fixed matrix registration when the test is intentionally helper-only or manually invoked; state that reason in the final response.
+
+### Step 4: Run Local Validation
 
 - Run the exact existing test files you changed, if any.
+- For new registered tests, run the same shape CI will use, for example `python tests/test_new_file.py`.
 - Run repository-wide checks only when they are already part of the task or workflow.
 - Avoid documenting placeholder test commands that may not exist in the current tree.
 
-### Step 4: Update Workflow Template Correctly
+### Step 5: Keep Workflow Template as Source of Truth
 
-For CI workflow changes:
+For CI workflow changes unrelated to a new, moved, or renamed test:
 
 1. Edit `.github/workflows/pr-test.yml.j2`
 2. Regenerate workflows:
@@ -59,18 +77,20 @@ python .github/workflows/generate_github_workflows.py
 
 3. Include both the template and generated workflow file in the change set (`.j2` and `.yml`). If the user asked for a commit, commit both.
 
-### Step 5: Provide Verifiable PR Notes
+### Step 6: Provide Verifiable PR Notes
 
 Include:
 
 - Which tests were added/changed
+- Where each new/renamed test was registered in `.github/workflows/pr-test.yml.j2`
 - Exact commands executed
 - GPU assumptions for each test path
 - Why this coverage protects against regression
 
 ## Common Mistakes
 
 - Editing generated workflow file only
+- Relying on `run-ci-changed` discovery for a new test that should run in the regular PR matrix
 - Forgetting `NUM_GPUS = 0` on a CPU-only changed test, causing `run-ci-changed` to default to 8 GPUs
 - Adding a CPU pytest file that passes under `pytest tests/foo.py` but fails under CI's `python tests/foo.py`
 - Adding tests without following existing constants/conventions

.claude/skills/vime-code-review-preferences/SKILL.md

@@ -0,0 +1,29 @@
+---
+name: vime-code-review-preferences
+description: Use when reviewing or editing vime code, especially refactors around helper APIs, branch selection, argument validation, or recurring reviewer preferences about avoiding unnecessary wrappers and making control flow self-explanatory.
+---
+
+# Vime Code Review Preferences
+
+Apply these lightweight review heuristics when changing vime code.
+
+## Prefer Direct APIs Over Thin Wrappers
+
+- Remove helper layers that only rename a call, format one path, or forward arguments without owning meaningful behavior.
+- Prefer calling the concrete reusable API directly, for example a `*_to_path` helper when the caller already knows the destination path.
+- Keep a wrapper only if it owns a real boundary: compatibility, validation, nontrivial error policy, lifecycle management, metrics/logging semantics, async/retry behavior, or cross-module ownership.
+- Avoid moving a redundant wrapper's body into another file just to preserve the wrapper shape. Inline the simple call at the natural ownership site.
+- When removing a wrapper, search for sibling wrappers and nearby helpers with `rg` and delete confirmed dead functions in the same pass.
+- Treat single-use convenience functions as suspicious when their only job is path formatting plus forwarding. Prefer the caller owning that one line.
+
+## Make Branches Explain Themselves
+
+- Order conditionals by semantic precedence: special transport/lifecycle modes first, then explicit mode choices, then default paths.
+- Prefer predicates that fully describe the branch, such as `mode == "full" and transport == "disk"`, over a broad predicate followed by an assert that explains what the branch really meant.
+- Use asserts as invariants for impossible states after validation, not as a substitute for clear branch conditions.
+
+## Keep Abstractions Honest
+
+- Add an abstraction only when it removes real duplication, hides fragile mechanics, or clarifies ownership.
+- When a review comment points out repeated indirection, look for a smaller public surface rather than adding another alias.
+- Preserve existing behavior intentionally. If cleanup changes error handling, logging, or failure visibility, call that out in the final response.

docker/Dockerfile

@@ -15,7 +15,7 @@ WORKDIR /root/
 # vllm/vllm-openai base is an inference image — add cu12 dev headers + cmake/git
 # so TE / apex / flash-attn source builds find cusparse.h etc.
 RUN apt-get update && apt-get install -y \
-    nvtop rsync dnsutils git cmake \
+    nvtop rsync dnsutils prometheus git cmake \
     cuda-nvrtc-dev-12-9 cuda-nvml-dev-12-9 cuda-profiler-api-12-9 cuda-nvtx-12-9 \
     libcusparse-dev-12-9 libcusolver-dev-12-9 libcufft-dev-12-9 libcurand-dev-12-9 \
     libcudnn9-dev-cuda-12 && \
@@ -42,6 +42,7 @@ RUN git clone https://github.com/Dao-AILab/flash-attention.git && \
 RUN pip install git+https://github.com/ISEEKYAN/mbridge.git@89eb10887887bc74853f89a4de258c0702932a1c --no-deps
 
 RUN pip install flash-linear-attention==0.4.1
+# FlashQLA: optional GDN backend for Qwen3.5/Qwen3-Next (--qwen-gdn-backend flashqla; requires SM90+)
 ARG INSTALL_FLASHQLA=0
 RUN if [ "${INSTALL_FLASHQLA}" = "1" ]; then \
       pip install git+https://github.com/QwenLM/FlashQLA.git --no-build-isolation; \
@@ -64,7 +65,8 @@ RUN git clone https://github.com/NVIDIA/Megatron-LM.git --recursive && \
     cd Megatron-LM && git checkout ${MEGATRON_COMMIT}
 
 # torch_memory_saver pinned to a193d9dd (upstream slime #1916).
-RUN pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@a193d9dd1b877d33c64a41cfb3db9f867df2d926 --no-cache-dir --force-reinstall
+# TMS_CUDA_MAJOR is required by this pin's build backend for CUDA wheels; base is cu129 -> 12.
+RUN TMS_CUDA_MAJOR=12 pip install git+https://github.com/fzyzcjy/torch_memory_saver.git@a193d9dd1b877d33c64a41cfb3db9f867df2d926 --no-cache-dir --force-reinstall
 RUN pip install git+https://github.com/radixark/Megatron-Bridge.git@bridge --no-deps --no-build-isolation
 RUN pip install nvidia-modelopt[torch]>=0.37.0 --no-build-isolation
 
@@ -75,21 +77,18 @@ RUN pip install --ignore-installed PyJWT && \
 # https://github.com/pytorch/pytorch/issues/168167
 RUN pip install nvidia-cudnn-cu12==9.16.0.29
 
-# reinstall numpy 1.x for megatron
-RUN pip install "numpy<2"
+# reinstall numpy 1.x for megatron; pin scipy<1.18 alongside it. vime's vllm/vllm-openai
+# base ships NO scipy, so unpinned it pulls scipy>=1.18, which hard-requires numpy>=2 and
+# uses np.long (removed in numpy>=1.24) -> AttributeError against the numpy<2 above ->
+# `import scipy/transformers` crash. slime's sglang base resolves scipy to 1.17.1 (numpy-1.x
+# compatible) natively, so slime needs no scipy pin -- this is a vime base-image divergence.
+# Real boundary is 1.18 (slime runs 1.17.1 fine), not the earlier 1.14 guess.
+RUN pip install "numpy<2" "scipy<1.18"
 
-RUN pip install IPython
-
-# Pin vllm-router explicitly so the vllm rollout routing layer is a visible build step
-# (also in requirements.txt; pulling it here makes the layer cache-able and fail-fast).
-RUN pip install "vllm-router>=0.1.14"
-
-RUN rm -rf /root/.cache/pip
+RUN rm -rf /root/.cache/pip /root/flash-attention
 
 # ====================================== Patches ============================================
 
-# Patch megatron BEFORE pip install -e . so any patch hunks that touch setup.py
-# or C++/CUDA extensions are picked up by the build.
 COPY docker/patch/${PATCH_VERSION}/megatron.patch /root/Megatron-LM/
 RUN cd Megatron-LM && \
     git update-index --refresh && \
@@ -123,21 +122,9 @@ RUN git clone https://github.com/vllm-project/vime.git /root/vime && \
 RUN cd /root/vime/vime/backends/megatron_utils/kernels/int4_qat && \
     pip install . --no-build-isolation
 
-# ====================================== Build-time smoke ============================================
-
-# Fail-fast import smoke + flashinfer version pin check. Catches ABI / version
-# regressions at build time instead of first GPU run.
-RUN python3 -c "\
-import vllm, vime, flashinfer; \
-from vime.backends.vllm_utils.vllm_engine import VLLMEngine; \
-print('vllm', vllm.__version__); \
-print('flashinfer', flashinfer.__version__); \
-print('VLLMEngine import ok')"
-
-# Megatron stream serialization. vllm rollout + megatron actor co-located on
-# the same GPU need this; default value causes stream-level contention.
-ENV CUDA_DEVICE_MAX_CONNECTIONS=1
-
-# Reset ENTRYPOINT inherited from vllm/vllm-openai base (`vllm serve`).
+# Reset ENTRYPOINT inherited from the vllm/vllm-openai base (`vllm serve`), so the
+# image is a plain bash/ray environment. Without this, `docker run ... bash -c ...`
+# and `ray job submit` append to `vllm serve` and break. Base-image-coupled: slime's
+# base has no such entrypoint, so slime's Dockerfile doesn't need this.
 ENTRYPOINT []
 CMD ["/bin/bash"]

docker/patch/latest/vllm.patch

@@ -1,3 +1,35 @@
+diff --git a/vllm/v1/engine/core.py b/vllm/v1/engine/core.py
+--- a/vllm/v1/engine/core.py
++++ b/vllm/v1/engine/core.py
+@@ -775,8 +775,10 @@
+         if tags is None or tags:
+             self.model_executor.wake_up(tags)
+
+-        # Resume scheduling (applies to all levels)
+-        self.resume_scheduler()
++        # Partial wakes intentionally keep the remaining allocations asleep.
++        # Resume scheduling only once all executor memory is resident again.
++        if not self.model_executor.is_sleeping:
++            self.resume_scheduler()
+
+     def is_sleeping(self) -> bool:
+         """Check if engine is sleeping at any level."""
+@@ -1894,9 +1896,12 @@
+                     continue
+
+                 # We are in a running state and so must execute a dummy pass
+-                # if the model didn't execute any ready requests.
+-                with self.log_iteration_details(None):
+-                    self.execute_dummy_batch()
++                # if the model didn't execute any ready requests -- unless the executor is
++                # asleep (#44483: a decode-shaped dummy batch reads freed KV -> illegal memory
++                # access). The finished-sync all-reduce below still runs (DP lockstep).
++                if not self.is_sleeping():
++                    with self.log_iteration_details(None):
++                        self.execute_dummy_batch()
+
+             # 3) All-reduce operation to determine global unfinished reqs.
+             self.engines_running = self._has_global_unfinished_reqs(
 diff --git a/vllm/model_executor/layers/fused_moe/all2all_utils.py b/vllm/model_executor/layers/fused_moe/all2all_utils.py
 --- a/vllm/model_executor/layers/fused_moe/all2all_utils.py
 +++ b/vllm/model_executor/layers/fused_moe/all2all_utils.py

docker/version.txt

@@ -1 +1 @@
-nightly-dev-20260519a
+nightly-dev-20260618a