sync: mirror Slime #2432 Straw replay and queue changes

.buildkite/README.md

@@ -17,9 +17,9 @@ The four test steps depend on the pre-commit gate. Each suite runs its files
 sequentially inside one step because these queues boot a fresh EC2 instance
 per job — a per-file matrix would be mostly boot + pip-install time.
 Most always-on CPU steps use the standard `python:3.11` image and install their
-lightweight dependencies at runtime. `upstream-sync-cpu` uses
-`vllm/vime:latest` because the synchronized GLM and checkpoint tests import the
-image-pinned Megatron stack even though they do not allocate a GPU.
+lightweight dependencies at runtime. `upstream-sync-cpu` uses `VIME_CI_IMAGE`
+(defaulting to `vllm/vime:latest`) because the synchronized GLM and checkpoint
+tests import the image-pinned Megatron stack even though they do not allocate a GPU.
 
 ## Creating the pipeline (one-time, Buildkite UI)
 
@@ -61,6 +61,10 @@ the per-test `VIME_TEST_USE_DEEPEP` / `VIME_TEST_USE_FP8_ROLLOUT` /
 The block uses `blocked_state: passed`, so a build whose CPU steps are green
 reports a passing commit status even if nobody unblocks the GPU gate.
 
+For sync validation, always select all six suites.  The REST unblock payload
+passes the selection directly, for example `{"gpu-suites": "short\nvllm-config\nmegatron\nvime-customized\nprecision\nckpt"}`;
+do not wrap it in a second `fields` object.
+
 GPU jobs run on the shared **`mithril-h100-pool`** queue, following the same
 pattern vllm-omni uses for it: each job is a Kubernetes pod (agent-stack-k8s
 `kubernetes` plugin) on an H100 SXM node, with GPUs allocated via
@@ -70,8 +74,8 @@ startup, so a warm HF cache is all they need. `WANDB_API_KEY` is not wired up
 yet; runs report without wandb until it's added (e.g. as a k8s secret in the
 pod spec).
 
-GPU jobs use `vllm/vime:latest`. Rebuild and publish that image before validating
-Dockerfile or vLLM patch changes.
+Set `VIME_CI_IMAGE` to an immutable candidate digest for image-backed jobs;
+otherwise they use `vllm/vime:latest`. Do not update `latest` before merge.
 
 ## Keeping it in sync
 

.buildkite/gpu_suites.py

@@ -23,15 +23,14 @@
 import subprocess
 
 GPU_QUEUE = "mithril-h100-pool"
-CI_IMAGE = "vllm/vime:latest"
+CI_IMAGE = os.environ.get("VIME_CI_IMAGE", "vllm/vime:latest")
 HF_CACHE_HOST_PATH = "/mnt/hf-cache"
 HF_HOME = "/root/.cache/huggingface"
 NODE_INSTANCE_TYPE = "gpu-h100-sxm"
 
 # (test_file, num_gpus, extra_args, env overrides)
 SUITES = {
     "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_fully_async_short.py", 4, "", {}),
     ],
@@ -64,8 +63,10 @@
         ("test_moonlight_16B_A3B_r3.py", 8, "", {"ENABLE_EVAL": "0"}),
         ("test_mimo_7B_mtp_only_grad.py", 8, "", {}),
         ("test_qwen2.5_0.5B_debug_rollout_then_train.py", 8, "", {}),
+        ("test_straw_checkpoint_fork.py", 4, "", {}),
         ("test_qwen2.5_0.5B_opd_vllm.py", 8, "", {}),
         ("test_qwen2.5_0.5B_fanout_short.py", 4, "", {}),
+        ("test_qwen2.5_0.5B_score_centering.py", 2, "", {}),
         ("test_qwen2.5_0.5B_debug_train_dump_e2e.py", 8, "", {}),
         ("test_qwen3_4B_external_pd.py", 6, "", {"VIME_TEST_UPDATE_MODE": "delta"}),
     ],

.buildkite/pipeline.yml

@@ -64,9 +64,10 @@ steps:
         python:3.11 bash -c '
           set -euo pipefail
           pip install -q torch --index-url https://download.pytorch.org/whl/cpu
-          pip install -q pytest numpy packaging pyyaml omegaconf tqdm httpx requests ray pybase64 pylatexenc sympy aiohttp pillow safetensors transformers cloudpickle blake3 xxhash zstandard psutil wandb
+          pip install -q pytest numpy packaging pyyaml omegaconf tqdm httpx requests ray pybase64 pylatexenc sympy aiohttp pillow safetensors transformers cloudpickle blake3 xxhash zstandard psutil wandb "straw-queue>=0.1.2"
           pip install -q -e . --no-deps
           python tests/test_megatron_argument_validation.py
+          python tests/test_optional_straw.py
           python tests/test_value_temperature.py
           python tests/test_docs_consistency.py
           python tests/test_placement_group.py
@@ -140,18 +141,26 @@ steps:
         -e GIT_CONFIG_PARAMETERS="'safe.directory=/workspace'" \
         -e GLOO_SOCKET_IFNAME=lo -e TP_SOCKET_IFNAME=lo \
         -v "$$PWD:/workspace" -w /workspace \
-        vllm/vime:latest bash -lc '
+        "$${VIME_CI_IMAGE:-vllm/vime:latest}" bash -lc '
           set -euo pipefail
+          pip install -q "straw-queue>=0.1.2" --break-system-packages
           pip install -q -e . --no-deps --break-system-packages
           for test_file in \
             tests/test_advantage_whiten_cp.py \
             tests/test_accelerator.py \
             tests/test_block_fp8_zero_block.py \
+            tests/test_cuda_stack_offload.py \
             tests/test_deep_ep_tms_patch.py \
+            tests/test_distributed_rollout.py \
             tests/test_discounted_returns.py \
             tests/test_eval_config.py \
             tests/test_filter_long_prompt.py \
             tests/test_fully_async_rollout.py \
+            tests/test_queue_sample_codec.py \
+            tests/test_rollout_straw_transport.py \
+            tests/test_routed_experts_disk_layout.py \
+            tests/test_straw_fully_async_recovery.py \
+            tests/test_straw_r3.py \
             tests/test_glm52_layerwise_comparison.py \
             tests/test_glm5_indexer_q_norm.py \
             tests/test_glm5_indexer_short_context.py \
@@ -160,11 +169,17 @@ steps:
             tests/test_model_provider_freeze.py \
             tests/test_policy_loss.py \
             tests/test_ppo_kl_metric.py \
+            tests/test_published_payload_retry.py \
+            tests/test_rollout_metadata_index.py \
+            tests/test_score_centering.py \
+            tests/test_score_centering_transport.py \
             tests/test_process_rollout_data.py \
             tests/test_qwen3_5_vl_native.py \
             tests/test_qwen3_linear_attention_cu_seqlens.py \
             tests/test_read_file_slicing.py \
             tests/test_reloadable_process_group_world.py \
+            tests/test_r3_spill.py \
+            tests/test_rollout_buffer_order.py \
             tests/test_rollout_data_utils.py \
             tests/test_rollout_metrics.py \
             tests/test_rollout_sample_hooks.py \
@@ -174,6 +189,7 @@ steps:
             tests/test_tau_bench_token_delta.py \
             tests/test_train_data_utils.py \
             tests/test_update_weight_factory.py \
+            tests/test_vllm_server_control.py \
             tests/observability/test_trace_utils.py; do
             python -m pytest "$$test_file"
           done
@@ -222,7 +238,7 @@ steps:
             value: short
           - label: "run-ci-vllm-config — 8 GPU, 4 tests"
             value: vllm-config
-          - label: "run-ci-megatron — up to 8 GPU, 21 runs"
+          - label: "run-ci-megatron — up to 8 GPU, 23 runs"
             value: megatron
           - label: "run-ci-vime-customized — 1–8 GPU, 5 tests"
             value: vime-customized

.claude/skills/add-dynamic-filter/SKILL.md

@@ -42,7 +42,7 @@ from vime.rollout.filter_hub.base_types import DynamicFilterOutput
 return DynamicFilterOutput(keep=True, reason=None)
 ```
 
-Buffer filter (called in `vime/rollout/data_source.py`):
+Buffer filter (called in `vime/data/data_source.py`):
 
 ```python
 def buffer_filter(args, rollout_id, buffer, num_samples):
@@ -98,4 +98,4 @@ Example wiring:
 - Dynamic filter types: `vime/rollout/filter_hub/base_types.py`
 - Dynamic filter example: `vime/rollout/filter_hub/dynamic_sampling_filters.py`
 - Rollout generation hook points: `vime/rollout/vllm_rollout.py`
-- Buffer filter hook point: `vime/rollout/data_source.py`
+- Buffer filter hook point: `vime/data/data_source.py`

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

@@ -27,3 +27,10 @@ Apply these lightweight review heuristics when changing vime code.
 - 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.
+
+## Keep Runtime State Out of Args
+
+- Treat `args` as configuration. Do not add ad hoc `args._xxx` attributes to pass actor handles, caches, recovery progress, current versions, or other runtime state between components. Renaming these to public attributes does not fix the ownership problem.
+- Put live state on the component that owns its lifecycle. Pass dependencies explicitly, and use typed return values for startup plans or restoration results shared between components.
+- Avoid replacing scattered temporary fields with a generic `args.context` bag. Keep each state object's responsibility and lifetime clear.
+- Preserve existing custom hook signatures, call arguments, and return conventions when refactoring internal state. User implementations must not need new parameters merely to accommodate an internal cleanup.

README.md

@@ -53,6 +53,8 @@ The vLLM community horizontally supports many LLM post-training frameworks, incl
 - **rollout (vLLM + router)**: Launches vLLM inference engines and routes generation requests; custom generate functions can wrap generation with multi-turn loops, tool calls, environment/sandbox interaction, and verifier-based rewards.
 - **data buffer**: A bridge module that manages prompt initialization, custom data, and rollout generation methods, including agentic workflows that produce samples through the same interface.
 
+The default payload transport is Ray `object-store`. `--rollout-data-transport straw` persists prompt tasks, rollout continuations, and training batches on shared storage. See the [straw architecture and recovery guide](docs/en/advanced/straw.md).
+
 ## Quick Start
 
 For a comprehensive quick start guide covering environment setup, data preparation, training startup, and key code analysis, please refer to:

README_zh.md

@@ -53,6 +53,8 @@ vLLM 社区横向支持许多 LLM post-training 框架，包括（按字母顺
 - **rollout (vLLM + router)**：启动 vLLM 推理引擎并路由生成请求；自定义生成函数可以在其上封装多轮循环、工具调用、环境/沙盒交互和基于 verifier 的奖励；
 - **data buffer**：桥梁模块，管理 prompt 初始化、自定义数据与 rollout 生成方法，包括通过同一接口产出样本的 agent 工作流。
 
+默认载荷传输为 Ray `object-store`。`--rollout-data-transport straw` 可在共享存储上持久化 prompt 任务、rollout continuation 和训练 batch。详见 [straw 架构与恢复指南](docs/zh/advanced/straw.md)。
+
 ## 快速开始
 
 有关环境配置、数据准备、训练启动和关键代码分析的完整快速开始指南，请参考：

docker/Dockerfile

@@ -153,15 +153,18 @@ RUN cd Megatron-LM && \
 # so apply the independently maintained patches in their validated order.
 COPY docker/patch/${PATCH_VERSION}/vllm-pull_weights.patch /tmp/vllm-pull_weights.patch
 COPY docker/patch/${PATCH_VERSION}/vllm.patch /tmp/vllm.patch
+COPY docker/patch/${PATCH_VERSION}/vllm-score-centering.patch /tmp/vllm-score-centering.patch
 COPY docker/patch/${PATCH_VERSION}/vllm-pd-request-metrics.patch /tmp/vllm-pd-request-metrics.patch
 COPY docker/patch/${PATCH_VERSION}/vllm-inflight-queue-diagnostics.patch /tmp/vllm-inflight-queue-diagnostics.patch
 RUN VLLM_SITE="$(python3 -c 'import os, vllm; print(os.path.dirname(os.path.dirname(vllm.__file__)))')" && \
     cd "$VLLM_SITE" && \
     git apply -v /tmp/vllm-pull_weights.patch && \
     git apply -v --allow-empty /tmp/vllm.patch && \
+    git apply -v /tmp/vllm-score-centering.patch && \
     git apply -v /tmp/vllm-pd-request-metrics.patch && \
     git apply -v /tmp/vllm-inflight-queue-diagnostics.patch && \
     rm /tmp/vllm-pull_weights.patch /tmp/vllm.patch \
+       /tmp/vllm-score-centering.patch \
        /tmp/vllm-pd-request-metrics.patch \
        /tmp/vllm-inflight-queue-diagnostics.patch
 

docker/patch/latest/megatron.patch

@@ -1005,3 +1005,36 @@ index e9736ac08..6567ed426 100644
                  )
                  for (model_chunk_idx, model_chunk) in enumerate(model)
              ]
+diff --git a/megatron/training/checkpointing.py b/megatron/training/checkpointing.py
+index 7c87eca191a525915502a1deb7622f03990c89de..bdb62df91b9ca2fe6e13d166a21f38e1b1e90e64 100644
+--- a/megatron/training/checkpointing.py
++++ b/megatron/training/checkpointing.py
+@@ -212,8 +212,8 @@
+             iteration, release = read_metadata(tracker_filename)
+
+     # Allow user to specify the loaded iteration.
+-    if getattr(args, "ckpt_step", None):
+-        iteration = args.ckpt_step
++    if getattr(args, "ckpt_step", None) is not None:
++        iteration, release = args.ckpt_step, False
+
+     return get_checkpoint_name(load_dir, iteration, release, return_base_dir=True)
+
+@@ -1185,14 +1185,14 @@
+             iteration, release = read_metadata(tracker_filename)
+
+     # Allow user to specify the loaded iteration.
+-    if getattr(args, "ckpt_step", None):
+-        iteration = args.ckpt_step
++    if getattr(args, "ckpt_step", None) is not None:
++        iteration, release = args.ckpt_step, False
+
+     # Record the iteration loaded (stored separately from args to avoid
+     # polluting checkpoints, since args is saved in checkpoints).
+     set_loaded_iteration(iteration)
+
+-    if non_persistent_iteration != -1:  # there is a non-persistent checkpoint
++    if non_persistent_iteration != -1 and getattr(args, "ckpt_step", None) is None:
+         if non_persistent_iteration >= iteration:
+             return _load_non_persistent_base_checkpoint(
+                 non_persistent_global_dir,