Feat: Support verl release/v0.8.0 and harden NPU vLLM workflows

.github/workflows/cpu_unit_tests.yml

@@ -31,10 +31,10 @@ jobs:
           cache: pip
           cache-dependency-path: ci/requirements-ci.txt
 
-      - name: Check out pinned upstream verl
+      - name: Check out release/v0.8.0 upstream verl
         run: |
           set -euo pipefail
-          verl_ref="$(awk -F= '/^VERL_BASE_COMMIT=/{print $2}' REQUIRED_VERL.txt)"
+          verl_ref="$(awk -F= '/^VERL_BASE_BRANCH=/{print $2}' REQUIRED_VERL.txt)"
           mkdir -p .ci
           git clone --filter=blob:none https://github.com/verl-project/verl .ci/upstream-verl
           git -C .ci/upstream-verl checkout "${verl_ref}"

Dockerfile.sglang

@@ -1,17 +1,17 @@
 # GPU SGLang runtime image.
 FROM verlai/verl:sgl0512.dev1
 
-ARG VERL_COMMIT=7aed6b230776f963fa09509c10d9c3a767d1102c
+ARG VERL_REF=release/v0.8.0
 ARG VERL_REPO=https://github.com/verl-project/verl.git
 
 WORKDIR /workspace
 
 RUN git clone ${VERL_REPO} /workspace/verl \
     && cd /workspace/verl \
-    && git checkout ${VERL_COMMIT} \
+    && git checkout ${VERL_REF} \
     && pip install -e .
 
 COPY . /workspace/verl-SpeCo
 
-ENV PYTHONPATH=/workspace/verl-SpeCo:${PYTHONPATH}
 WORKDIR /workspace/verl-SpeCo
+RUN pip install -e .

Dockerfile.vllm

@@ -1,17 +1,17 @@
 # GPU vLLM runtime image.
 FROM verlai/verl:vllm023.dev1
 
-ARG VERL_COMMIT=7aed6b230776f963fa09509c10d9c3a767d1102c
+ARG VERL_REF=release/v0.8.0
 ARG VERL_REPO=https://github.com/verl-project/verl.git
 
 WORKDIR /workspace
 
 RUN git clone ${VERL_REPO} /workspace/verl \
     && cd /workspace/verl \
-    && git checkout ${VERL_COMMIT} \
+    && git checkout ${VERL_REF} \
     && pip install -e .
 
 COPY . /workspace/verl-SpeCo
 
-ENV PYTHONPATH=/workspace/verl-SpeCo:${PYTHONPATH}
 WORKDIR /workspace/verl-SpeCo
+RUN pip install -e .

README.md

@@ -103,31 +103,38 @@ ci/                               # smoke-test helpers and CI notes
 
 ## Installation
 
-This repository does not currently define its own Python package metadata. Use
-it with the upstream `verl` commit pinned in
+Install the upstream `verl` release branch specified in
 [`REQUIRED_VERL.txt`](./REQUIRED_VERL.txt), which is mirrored in
 [`verl_speco/config/speco_base.yaml`](./verl_speco/config/speco_base.yaml).
 By default, unsupported `verl` versions produce a warning. Set
 `VERL_SPECO_STRICT_VERL=1` to fail closed when the importable `verl` does not
-match the pinned version or commit.
+match the release/v0.8.0 version and API contract.
 
 One typical editable setup is:
 
 ```bash
 git clone https://github.com/verl-project/verl.git
 cd verl
-git checkout 7aed6b230776f963fa09509c10d9c3a767d1102c
+git checkout release/v0.8.0
 pip install -e .
 
-cd /path/to/verl-SpeCo
-export PYTHONPATH="$PWD:$PYTHONPATH"
+cd ..
+git clone https://github.com/verl-project/verl-SpeCo.git
+cd verl-SpeCo
+pip install -e .
 ```
 
+The editable install exposes the `verl_speco` package without modifying
+`PYTHONPATH`. It also installs the `verl-speco`, `verl-speco-draft-train`, and
+`verl-speco-inspect-features` command-line entry points. Install the matching
+GPU or NPU rollout runtime separately; `verl-SpeCo` intentionally does not let
+pip replace accelerator-specific PyTorch, vLLM, SGLang, or vLLM-Ascend builds.
+
 ### Docker Images
 
 You can also build GPU runtime images from the official `verlai/verl`
-development images and then pin the importable upstream `verl` checkout to the
-required v0.8.0 commit. The Dockerfiles below target GPU deployments; use the
+development images and then use the importable upstream `verl` checkout from
+the release/v0.8.0 branch. The Dockerfiles below target GPU deployments; use the
 matching accelerator image for NPU or other accelerator runtimes.
 
 For GPU vLLM-based examples, use this Dockerfile:
@@ -136,20 +143,20 @@ For GPU vLLM-based examples, use this Dockerfile:
 # GPU vLLM runtime image.
 FROM verlai/verl:vllm023.dev1
 
-ARG VERL_COMMIT=7aed6b230776f963fa09509c10d9c3a767d1102c
+ARG VERL_REF=release/v0.8.0
 ARG VERL_REPO=https://github.com/verl-project/verl.git
 
 WORKDIR /workspace
 
 RUN git clone ${VERL_REPO} /workspace/verl \
     && cd /workspace/verl \
-    && git checkout ${VERL_COMMIT} \
+    && git checkout ${VERL_REF} \
     && pip install -e .
 
 COPY . /workspace/verl-SpeCo
 
-ENV PYTHONPATH=/workspace/verl-SpeCo:${PYTHONPATH}
 WORKDIR /workspace/verl-SpeCo
+RUN pip install -e .
 ```
 
 Build it from the `verl-SpeCo` repository root:
@@ -164,20 +171,20 @@ For GPU SGLang-based examples, use the same layout with the SGLang base image:
 # GPU SGLang runtime image.
 FROM verlai/verl:sgl0512.dev1
 
-ARG VERL_COMMIT=7aed6b230776f963fa09509c10d9c3a767d1102c
+ARG VERL_REF=release/v0.8.0
 ARG VERL_REPO=https://github.com/verl-project/verl.git
 
 WORKDIR /workspace
 
 RUN git clone ${VERL_REPO} /workspace/verl \
     && cd /workspace/verl \
-    && git checkout ${VERL_COMMIT} \
+    && git checkout ${VERL_REF} \
     && pip install -e .
 
 COPY . /workspace/verl-SpeCo
 
-ENV PYTHONPATH=/workspace/verl-SpeCo:${PYTHONPATH}
 WORKDIR /workspace/verl-SpeCo
+RUN pip install -e .
 ```
 
 Build it from the `verl-SpeCo` repository root:
@@ -289,7 +296,7 @@ pip install -r ci/requirements-ci.txt
 pytest tests
 ```
 
-Some tests require a pinned upstream `verl` checkout. Set
+Some tests require an upstream `verl` checkout from `release/v0.8.0`. Set
 `VERL_SPECO_UPSTREAM_ROOT` to the root of that checkout when running the config
 composition contract:
 

REQUIRED_VERL.txt

@@ -1,5 +1,5 @@
-VERL_BASE_TAG=v0.8.0
-VERL_BASE_COMMIT=7aed6b230776f963fa09509c10d9c3a767d1102c
+VERL_BASE_BRANCH=release/v0.8.0
+VERL_BASE_VERSION=0.8.0
 VERL_SOURCE_MODIFICATIONS_ALLOWED=false
 REQUIRES=import verl only
 COMPATIBILITY_POLICY=warn by default

examples/run_qwen3-8b_drafter_dflash_sglang.sh

@@ -5,6 +5,8 @@ exp_name='qwen3_8b_dflash_drafter'
 
 gen_tp=2
 train_sp=1
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -14,6 +16,9 @@ DRAFTER_PATH=/path/to/drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=64 \

examples/run_qwen3-8b_drafter_dflash_vllm.sh

@@ -6,6 +6,8 @@ exp_name='qwen3_8b_dflash_drafter_vllm_gpu'
 
 gen_tp=2
 train_sp=1
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -16,6 +18,9 @@ DRAFTER_PATH=/path/to/vllm-compatible-dflash-drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=16 \
@@ -91,7 +96,6 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     trainer.experiment_name=${exp_name} \
     trainer.n_gpus_per_node=16 \
     trainer.nnodes=1 \
-    trainer.resume_mode=disable \
     trainer.default_local_dir=${CKPTS_DIR} \
     trainer.total_training_steps=100 \
     trainer.save_freq=20 \

examples/run_qwen3-8b_drafter_dflash_vllm_npu.sh

@@ -1,5 +1,19 @@
 set -x
 export ASCEND_RT_VISIBLE_DEVICES="${ASCEND_RT_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
+case "${LD_PRELOAD:-}" in
+    *libjemalloc*) ;;
+    *)
+        if [ -f /usr/lib/aarch64-linux-gnu/libjemalloc.so.2 ]; then
+            export LD_PRELOAD="/usr/lib/aarch64-linux-gnu/libjemalloc.so.2${LD_PRELOAD:+:$LD_PRELOAD}"
+        elif [ -f /usr/lib64/libjemalloc.so.2 ]; then
+            export LD_PRELOAD="/usr/lib64/libjemalloc.so.2${LD_PRELOAD:+:$LD_PRELOAD}"
+        fi
+        ;;
+esac
+export MALLOC_CONF="${MALLOC_CONF:-narenas:8,thp:never,metadata_thp:disabled,dirty_decay_ms:0,muzzy_decay_ms:0}"
+export SPECO_JEMALLOC_RECLAIM_MODE="${SPECO_JEMALLOC_RECLAIM_MODE:-purge}"
+export MALLOC_ARENA_MAX="${MALLOC_ARENA_MAX:-2}"
+export MALLOC_TRIM_THRESHOLD_="${MALLOC_TRIM_THRESHOLD_:-131072}"
 
 # NPU example for the native DFlash proposer in vLLM.
 project_name='verl_grpo_example_dflash_drafter'
@@ -8,6 +22,8 @@ exp_name='qwen3_8b_dflash_drafter_vllm_npu'
 gen_tp=2
 train_sp=4
 ppo_gpus_per_node=${SPECO_ACCELERATOR_COUNT:-8}
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -18,6 +34,10 @@ DRAFTER_PATH=/path/to/vllm-compatible-dflash-drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    transfer_queue.enable=False \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=64 \
@@ -38,8 +58,8 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     actor_rollout_ref.actor.entropy_coeff=0 \
     actor_rollout_ref.actor.calculate_entropy=False \
     actor_rollout_ref.model.enable_gradient_checkpointing=True \
-    actor_rollout_ref.actor.fsdp_config.param_offload=True \
-    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
     actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=10 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=${gen_tp} \
     actor_rollout_ref.actor.ulysses_sequence_parallel_size=${train_sp} \
@@ -59,7 +79,7 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
     actor_rollout_ref.rollout.n=5 \
     actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=10 \
-    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    actor_rollout_ref.ref.fsdp_config.param_offload=False \
     actor_rollout_ref.rollout.drafter.enable=True \
     actor_rollout_ref.rollout.drafter.enable_drafter_training=True \
     actor_rollout_ref.rollout.drafter.model_path=${DRAFTER_PATH} \
@@ -97,7 +117,6 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     trainer.experiment_name=${exp_name} \
     trainer.n_gpus_per_node=${ppo_gpus_per_node} \
     trainer.nnodes=1 \
-    trainer.resume_mode=disable \
     trainer.default_local_dir=${CKPTS_DIR} \
     trainer.total_training_steps=200 \
     trainer.save_freq=20 \

examples/run_qwen3-8b_drafter_dspark_vllm.sh

@@ -7,6 +7,8 @@ exp_name='qwen3_8b_dspark_drafter_vllm_gpu'
 
 gen_tp=2
 train_sp=1
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -17,6 +19,9 @@ DRAFTER_PATH=/path/to/vllm-compatible-dspark-drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=16 \
@@ -102,7 +107,6 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     trainer.experiment_name=${exp_name} \
     trainer.n_gpus_per_node=16 \
     trainer.nnodes=1 \
-    trainer.resume_mode=disable \
     trainer.default_local_dir=${CKPTS_DIR} \
     trainer.total_training_steps=100 \
     trainer.save_freq=20 \

examples/run_qwen3-8b_drafter_dspark_vllm_npu.sh

@@ -1,5 +1,19 @@
 set -x
 export ASCEND_RT_VISIBLE_DEVICES="${ASCEND_RT_VISIBLE_DEVICES:-0,1,2,3,4,5,6,7}"
+case "${LD_PRELOAD:-}" in
+    *libjemalloc*) ;;
+    *)
+        if [ -f /usr/lib/aarch64-linux-gnu/libjemalloc.so.2 ]; then
+            export LD_PRELOAD="/usr/lib/aarch64-linux-gnu/libjemalloc.so.2${LD_PRELOAD:+:$LD_PRELOAD}"
+        elif [ -f /usr/lib64/libjemalloc.so.2 ]; then
+            export LD_PRELOAD="/usr/lib64/libjemalloc.so.2${LD_PRELOAD:+:$LD_PRELOAD}"
+        fi
+        ;;
+esac
+export MALLOC_CONF="${MALLOC_CONF:-narenas:8,thp:never,metadata_thp:disabled,dirty_decay_ms:0,muzzy_decay_ms:0}"
+export SPECO_JEMALLOC_RECLAIM_MODE="${SPECO_JEMALLOC_RECLAIM_MODE:-purge}"
+export MALLOC_ARENA_MAX="${MALLOC_ARENA_MAX:-2}"
+export MALLOC_TRIM_THRESHOLD_="${MALLOC_TRIM_THRESHOLD_:-131072}"
 
 # NPU example for DSpark on vLLM-Ascend. SPECO keeps the user-facing
 # algorithm as DSPARK and maps it to vLLM's dflash speculative method.
@@ -9,6 +23,8 @@ exp_name='qwen3_8b_dspark_drafter_vllm_npu'
 gen_tp=2
 train_sp=4
 ppo_gpus_per_node=${SPECO_ACCELERATOR_COUNT:-8}
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -19,6 +35,10 @@ DRAFTER_PATH=/path/to/vllm-compatible-dspark-drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    transfer_queue.enable=False \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=64 \
@@ -39,8 +59,8 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     actor_rollout_ref.actor.entropy_coeff=0 \
     actor_rollout_ref.actor.calculate_entropy=False \
     actor_rollout_ref.model.enable_gradient_checkpointing=True \
-    actor_rollout_ref.actor.fsdp_config.param_offload=True \
-    actor_rollout_ref.actor.fsdp_config.optimizer_offload=True \
+    actor_rollout_ref.actor.fsdp_config.param_offload=False \
+    actor_rollout_ref.actor.fsdp_config.optimizer_offload=False \
     actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=10 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=${gen_tp} \
     actor_rollout_ref.actor.ulysses_sequence_parallel_size=${train_sp} \
@@ -60,7 +80,7 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
     actor_rollout_ref.rollout.n=5 \
     actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=10 \
-    actor_rollout_ref.ref.fsdp_config.param_offload=True \
+    actor_rollout_ref.ref.fsdp_config.param_offload=False \
     actor_rollout_ref.rollout.drafter.enable=True \
     actor_rollout_ref.rollout.drafter.enable_drafter_training=True \
     actor_rollout_ref.rollout.drafter.model_path=${DRAFTER_PATH} \
@@ -105,7 +125,6 @@ PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     trainer.experiment_name=${exp_name} \
     trainer.n_gpus_per_node=${ppo_gpus_per_node} \
     trainer.nnodes=1 \
-    trainer.resume_mode=disable \
     trainer.default_local_dir=${CKPTS_DIR} \
     trainer.total_training_steps=200 \
     trainer.save_freq=20 \

examples/run_qwen3-8b_drafter_eagle3_sglang.sh

@@ -5,6 +5,8 @@ exp_name='qwen3_8b_function_rm_drafter'
 
 gen_tp=2
 train_sp=4
+ray_num_cpus=${SPECO_RAY_NUM_CPUS:-64}
+ray_worker_soft_limit=${SPECO_RAY_WORKER_SOFT_LIMIT:-8}
 
 MODEL_PATH=/path/to/model
 CKPTS_DIR=/path/to/checkpoint
@@ -14,6 +16,9 @@ DRAFTER_PATH=/path/to/drafter
 
 PYTHONUNBUFFERED=1 python3 -m verl_speco.main \
     algorithm.adv_estimator=grpo \
+    ray_kwargs.ray_init.num_cpus=${ray_num_cpus} \
+    +ray_kwargs.ray_init._system_config.prestart_worker_first_driver=false \
+    +ray_kwargs.ray_init._system_config.num_workers_soft_limit=${ray_worker_soft_limit} \
     data.train_files=${TRAIN_FILE} \
     data.val_files=${TEST_FILE} \
     data.train_batch_size=64 \