[sglang] feat: retires sglang spmd mode in the codebase

.github/workflows/sgl.yml

@@ -124,23 +124,6 @@ jobs:
       - name: Test the latest SGLang Rollout async with agent loop
         run: |
           ROLLOUT_NAME=sglang pytest -svvv tests/experimental/agent_loop
-      - name: Test the latest SGLang
-        run: |
-          cd tests/workers/rollout
-          torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_spmd.py
-      - name: Test the latest SGLang Rollout async with interaction
-        run: |
-          cd tests/workers/rollout
-          torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_async_rollout_w_interaction.py
-      - name: Test the latest SGLang Multi Interaction
-        run: |
-          cd tests/workers/rollout
-          torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_multi_interaction.py
-      - name: Test the latest SGLang Rollout async with tool
-        run: |
-          cd tests/workers/rollout
-          torchrun --nnodes=1 --nproc_per_node=2 $(which pytest) -s test_sglang_async_rollout_w_tools.py
-      # Note(haibin.lin): for any new test, please update gpu_unit_tests.yaml to avoid repeated tests
 
   cleanup:
     runs-on: ubuntu-latest

examples/grpo_trainer/run_gptoss_20b.sh

@@ -60,7 +60,7 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.rollout.log_prob_micro_batch_size_per_gpu=32 \
     actor_rollout_ref.rollout.tensor_model_parallel_size=2 \
     actor_rollout_ref.rollout.name=sglang \
-    actor_rollout_ref.rollout.mode=sync \
+    actor_rollout_ref.rollout.mode=async \
     actor_rollout_ref.rollout.engine_kwargs.sglang.attention_backend=triton \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.7 \
     actor_rollout_ref.rollout.n=5 \

examples/sglang_multiturn/run_qwen2.5-3b_gsm8k_multiturn.sh

@@ -49,7 +49,7 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
     actor_rollout_ref.ref.fsdp_config.param_offload=True \
     actor_rollout_ref.rollout.over_sample_rate=0.1 \
-    actor_rollout_ref.rollout.mode=sync \
+    actor_rollout_ref.rollout.mode=async \
     algorithm.use_kl_in_reward=False \
     trainer.critic_warmup=0 \
     trainer.logger='["console","wandb"]' \

examples/sglang_multiturn/run_qwen3-4b_gsm8k_multiturn.sh

@@ -36,7 +36,7 @@ python3 -m verl.trainer.main_ppo \
     actor_rollout_ref.rollout.gpu_memory_utilization=0.5 \
     actor_rollout_ref.rollout.n=16 \
     actor_rollout_ref.rollout.over_sample_rate=0.1 \
-    actor_rollout_ref.rollout.mode=sync \
+    actor_rollout_ref.rollout.mode=async \
     actor_rollout_ref.ref.log_prob_micro_batch_size_per_gpu=32 \
     actor_rollout_ref.ref.fsdp_config.param_offload=True \
     algorithm.use_kl_in_reward=False \

pyproject.toml

@@ -30,7 +30,7 @@ requires-python = ">=3.10"
 # Note: While the formatter will attempt to format lines such that they remain within the line-length,
 # it isn't a hard upper bound, and formatted lines may exceed the line-length.
 line-length = 120
-exclude = ["tests/workers/rollout/test_sglang_async_rollout_sf_tools.py", "scripts/legacy_model_merger.py"]
+exclude = ["scripts/legacy_model_merger.py"]
 
 [tool.ruff.lint]
 isort = {known-first-party = ["verl"]}