Zhichenzzz · Sep 21, 2026 · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026 · Sep 19, 2026
diff --git a/docker/Dockerfile b/docker/Dockerfile
@@ -11,12 +11,12 @@
 # release for the current arch picked from TARGETARCH — WHEELS_TAG_X86 on amd64,
 # WHEELS_TAG_ARM64 on arm64 (cu12-x86 overrides WHEELS_TAG_X86).
 
-ARG SGLANG_IMAGE_TAG=v0.5.19
+ARG SGLANG_IMAGE_TAG=v0.5.20
 FROM lmsysorg/sglang:${SGLANG_IMAGE_TAG} AS sglang
 
 # ======================================== Arguments =============================================
 
-ARG SGLANG_BRANCH=sglang-miles
+ARG SGLANG_BRANCH=sglang-miles-v0.5.20
 ARG SGLANG_COMMIT=""
 
 ARG MEGATRON_REPO=radixark/Megatron-LM

diff --git a/docker/build.py b/docker/build.py
@@ -47,6 +47,9 @@
         "build_args": {
             "ENABLE_CUDA_13": "0",
             "SGLANG_IMAGE_TAG": "v0.5.19-cu129",
+            # Frozen with the base image: sglang publishes no cu12 image from v0.5.20 on,
+            # so this variant cannot follow the branch the cu13 images track.
+            "SGLANG_BRANCH": "sglang-miles-v0.5.19-final",
             "WHEELS_TAG_X86": "cu129-x86_64",
         },
     },

diff --git a/docs/models/deepseek/deepseek-v3-2.md b/docs/models/deepseek/deepseek-v3-2.md
@@ -163,16 +163,16 @@ The `--enable-mis` flag turns on truncated importance sampling to correct the tr
 ```bash
 --sglang-mem-fraction-static 0.8
 --sglang-attention-backend nsa
---sglang-nsa-decode-backend flashmla_sparse
---sglang-nsa-prefill-backend flashmla_sparse
+--sglang-dsa-decode-backend flashmla_sparse
+--sglang-dsa-prefill-backend flashmla_sparse
 --sglang-kv-cache-dtype bf16
 --sglang-page-size 64  # the NSA KV cache requires 64 on CUDA
 --rollout-num-gpus-per-engine 8  # 2 with --use-single-node
 --sglang-tp-size 8
 --sglang-dp-size 8
 --sglang-enable-dp-attention
 --sglang-enable-dp-lm-head
---sglang-cuda-graph-max-bs 256
+--sglang-cuda-graph-max-bs-decode 256
 --sglang-moe-runner-backend flashinfer_trtllm_routed  # triton on H100 / H200
 ```
 

diff --git a/docs/models/deepseek/deepseek-v4-pro.md b/docs/models/deepseek/deepseek-v4-pro.md
@@ -85,7 +85,7 @@ SGLANG_ARGS=(
    --sglang-chunked-prefill-size 8192
    --sglang-server-concurrency 1024
    --sglang-moe-a2a-backend deepep             # DeepEP normal-mode dispatch
-   --sglang-cuda-graph-max-bs 8                # see hang caveat below
+   --sglang-cuda-graph-max-bs-decode 8                # see hang caveat below
    --sglang-mem-fraction-static 0.7            # Pro needs a larger dynamic buffer
 )
 ```
@@ -98,7 +98,7 @@ Megatron side: `--qkv-format bshd` (V4 needs `bshd` with CP-aware data slicing).
 
 1. **Engine size ≥ 32 GPUs.** Pro needs a single SGLang engine spanning at least 32 GPUs — the launcher hard-codes `--rollout-num-gpus-per-engine 32`. Smaller engines do not leave enough memory after weights, KV cache, indexer state, and DeepEP buffers, and rollout will OOM under load.
 2. **EP is mandatory; pure TP will not shard the model.** 384 routed experts × `moe_ffn_hidden_size=3072` cannot be partitioned by tensor parallelism alone — the model must use expert parallelism (`--sglang-ep-size 32`) to spread the expert MLPs across ranks. `--sglang-tp-size 32` only covers the attention / embedding paths.
-3. **DeepEP normal-mode + CUDA graphs can hang at large batch sizes.** When `--sglang-moe-a2a-backend deepep` is on, an overly large `--sglang-cuda-graph-max-bs` makes SGLang hang during graph capture or replay. The launcher pins it to `8` for Pro — raise it only after verifying the engine doesn't deadlock at your target batch.
+3. **DeepEP normal-mode + CUDA graphs can hang at large batch sizes.** When `--sglang-moe-a2a-backend deepep` is on, an overly large `--sglang-cuda-graph-max-bs-decode` makes SGLang hang during graph capture or replay. The launcher pins it to `8` for Pro — raise it only after verifying the engine doesn't deadlock at your target batch.
 
 </Warning>
 

diff --git a/docs/models/deepseek/deepseek.md b/docs/models/deepseek/deepseek.md
@@ -201,14 +201,14 @@ with attention DP 1 and 8 respectively.
 --sglang-server-concurrency 1024
 --sglang-max-running-requests 2048
 --sglang-chunked-prefill-size 16384
---sglang-cuda-graph-max-bs 256
+--sglang-cuda-graph-max-bs-decode 256
 ```
 
 `--rollout-num-gpus-per-engine` corresponds to SGLang's `tp_size`. To exploit large-EP inference,
 the recipe sets EP64, DP-attention with DP8, and DeepEP `low_latency`.
 `--sglang-server-concurrency` is a miles-specific knob to keep the SGLang HTTP server from being
 swamped — default 512, raised to 1024 here so each of the 8 DP ranks gets 128 concurrent requests.
-`SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK` is exported to match `--sglang-cuda-graph-max-bs`.
+`SGLANG_DEEPEP_NUM_MAX_DISPATCH_TOKENS_PER_RANK` is exported to match `--sglang-cuda-graph-max-bs-decode`.
 
 ### 5.4 Optimizer
 

diff --git a/docs/models/glm/glm5-2.md b/docs/models/glm/glm5-2.md
@@ -112,8 +112,8 @@ Always-on flags:
 
 # DSA / NSA attention with FP8 KV cache
 --sglang-kv-cache-dtype fp8_e4m3
---sglang-nsa-decode-backend flashmla_kv
---sglang-nsa-prefill-backend flashmla_sparse
+--sglang-dsa-decode-backend flashmla_kv
+--sglang-dsa-prefill-backend flashmla_sparse
 --sglang-attention-backend nsa
 --sglang-page-size 64
 

diff --git a/docs/models/glm/glm5.md b/docs/models/glm/glm5.md
@@ -89,8 +89,8 @@ Always-on flags:
 
 # DSA / NSA attention
 --sglang-page-size 64
---sglang-nsa-decode-backend flashmla_sparse
---sglang-nsa-prefill-backend flashmla_sparse
+--sglang-dsa-decode-backend flashmla_sparse
+--sglang-dsa-prefill-backend flashmla_sparse
 --sglang-attention-backend nsa
 
 --sglang-max-running-requests 512

diff --git a/docs/models/kimi/kimi-k2.5.md b/docs/models/kimi/kimi-k2.5.md
@@ -162,7 +162,7 @@ SGLANG_ARGS=(
    --sglang-mem-fraction-static 0.7
    --sglang-ep-size 8
    --sglang-server-concurrency 1024
-   --sglang-cuda-graph-bs 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
    --use-rollout-routing-replay
 )
 ```

diff --git a/docs/models/qwen/qwen3-5-moe.md b/docs/models/qwen/qwen3-5-moe.md
@@ -84,7 +84,7 @@ GRPO with `--eps-clip 0.2 --eps-clip-high 0.28 --use-kl-loss --kl-loss-coef 0.00
 --rollout-num-gpus-per-engine 8
 --sglang-mem-fraction-static 0.7
 --sglang-ep-size 8
---sglang-cuda-graph-bs 1 2 4 8 16 24 ... 256
+--sglang-cuda-graph-bs-decode 1 2 4 8 16 24 ... 256
 
 # mtp speculative decoding
 --sglang-speculative-algorithm EAGLE

diff --git a/docs/models/qwen/qwen3-6-moe.md b/docs/models/qwen/qwen3-6-moe.md
@@ -118,7 +118,7 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8
    --sglang-mem-fraction-static 0.7
    --sglang-ep-size 8
-   --sglang-cuda-graph-bs 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
 
    # MTP speculative decoding
    --sglang-speculative-algorithm EAGLE
@@ -127,7 +127,7 @@ SGLANG_ARGS=(
    --sglang-speculative-num-draft-tokens 3
 
    --sglang-max-running-requests 256
-   --sglang-mamba-scheduler-strategy extra_buffer
+   --sglang-mamba-radix-cache-strategy extra_buffer
 )
 ```
 

diff --git a/docs/models/qwen/qwen3-moe.md b/docs/models/qwen/qwen3-moe.md
@@ -99,7 +99,7 @@ python scripts/run_qwen3_235b_a22b.py
 ```bash
 --rollout-num-gpus-per-engine 8
 --sglang-mem-fraction-static 0.7
---sglang-cuda-graph-max-bs 512
+--sglang-cuda-graph-max-bs-decode 512
 ```
 
 `run_qwen3_235b_a22b.py`:
@@ -111,7 +111,7 @@ python scripts/run_qwen3_235b_a22b.py
 --sglang-dp-size 4
 --sglang-ep-size 32
 --sglang-enable-dp-lm-head
---sglang-cuda-graph-bs 1 2 4 8 16 24 ... 256
+--sglang-cuda-graph-bs-decode 1 2 4 8 16 24 ... 256
 --sglang-moe-a2a-backend deepep
 --sglang-deepep-mode auto
 ```

diff --git a/docs/models/qwen/qwen3-next.md b/docs/models/qwen/qwen3-next.md
@@ -92,7 +92,7 @@ Both topologies use GSPO (`--advantage-estimator gspo --eps-clip 4e-4`); `--use-
 --rollout-num-gpus-per-engine 8
 --sglang-mem-fraction-static 0.8
 --sglang-ep-size 8
---sglang-cuda-graph-bs 1 2 4 8 16 24 ... 128
+--sglang-cuda-graph-bs-decode 1 2 4 8 16 24 ... 128
 
 --sglang-speculative-algorithm EAGLE
 --sglang-speculative-num-steps 2

diff --git a/docs/models/qwen/qwen3.md b/docs/models/qwen/qwen3.md
@@ -102,7 +102,7 @@ Rollout uses `--rm-type deepscaler` against `dapo-math-17k`. The SFT recipe (`py
 --sglang-mem-fraction-static 0.7
 ```
 
-`Qwen3-32B` additionally pins `--sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)`. The FSDP variant uses `--attn-implementation flash_attention_3`, SGLang attention backend `fa3`, and adds `--update-weight-buffer-size 536870912 --gradient-checkpointing`.
+`Qwen3-32B` additionally pins `--sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)`. The FSDP variant uses `--attn-implementation flash_attention_3`, SGLang attention backend `fa3`, and adds `--update-weight-buffer-size 536870912 --gradient-checkpointing`.
 
 ### 5.4 Optimizer
 

diff --git a/docs/models/thinkingmachines/inkling.md b/docs/models/thinkingmachines/inkling.md
@@ -139,7 +139,7 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 16
    --sglang-attention-backend fa4
    --sglang-moe-runner-backend triton
-   --sglang-mamba-scheduler-strategy extra_buffer
+   --sglang-mamba-radix-cache-strategy extra_buffer
    --sglang-enable-multimodal
    --sglang-context-length 8192
 

diff --git a/examples/experimental/eval/scripts/run-qwen3-32B.sh b/examples/experimental/eval/scripts/run-qwen3-32B.sh
@@ -116,7 +116,7 @@ WANDB_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8
    --sglang-mem-fraction-static 0.7
-   # --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
+   # --sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)
 )
 
 MISC_ARGS=(

diff --git a/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh b/examples/experimental/multi_agent/run-qwen3-30B-A3B-multi-agent.sh
@@ -113,7 +113,7 @@ WANDB_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8
    --sglang-mem-fraction-static 0.7
-   --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
+   --sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)
 )
 
 MISC_ARGS=(

diff --git a/examples/experimental/openenv/glm52_tbench2/run_glm5_2_744b_a40b_daytona.py b/examples/experimental/openenv/glm52_tbench2/run_glm5_2_744b_a40b_daytona.py
@@ -229,11 +229,11 @@ def _execute_train(args: ScriptArgs):
         "--sglang-mem-fraction-static 0.85 "
         f"--sglang-ep-size {sglang_world_size} "
         "--sglang-kv-cache-dtype fp8_e4m3 "
-        "--sglang-nsa-decode-backend flashmla_kv "
-        "--sglang-nsa-prefill-backend flashmla_sparse "
+        "--sglang-dsa-decode-backend flashmla_kv "
+        "--sglang-dsa-prefill-backend flashmla_sparse "
         "--sglang-attention-backend nsa "
         "--sglang-page-size 64 "
-        "--sglang-cuda-graph-max-bs 32 "
+        "--sglang-cuda-graph-max-bs-decode 32 "
         f"--sglang-max-running-requests {256 if balanced else 512} "
         f"--sglang-chunked-prefill-size {32768 if balanced else 2048 * sglang_world_size} "
         "--sglang-watchdog-timeout 3600 "

diff --git a/examples/geo3k_vlm/multi_turn/run_geo3k_vlm_multi_turn.py b/examples/geo3k_vlm/multi_turn/run_geo3k_vlm_multi_turn.py
@@ -101,7 +101,7 @@ def execute():
     sglang_args = (
         "--rollout-num-gpus-per-engine 1 "
         "--sglang-mem-fraction-static 0.6 "
-        f"--sglang-cuda-graph-bs {' '.join(map(str, [1, 2, 4, 8] + list(range(16, 257, 8))))} "
+        f"--sglang-cuda-graph-bs-decode {' '.join(map(str, [1, 2, 4, 8] + list(range(16, 257, 8))))} "
     )
 
     fsdp_args = (

diff --git a/examples/geo3k_vlm/run_geo3k_vlm.sh b/examples/geo3k_vlm/run_geo3k_vlm.sh
@@ -130,7 +130,7 @@ OPTIMIZER_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 1
    --sglang-mem-fraction-static 0.6
-   --sglang-cuda-graph-bs 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 136 144 152 160 168 176 184 192 200 208 216 224 232 240 248 256
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128 136 144 152 160 168 176 184 192 200 208 216 224 232 240 248 256
 )
 
 # Wandb args (only if WANDB_API_KEY is set)

diff --git a/examples/infra_features/fully_async/run_qwen3_30b_a3b_fully_async.py b/examples/infra_features/fully_async/run_qwen3_30b_a3b_fully_async.py
@@ -123,7 +123,7 @@ def execute(args: ScriptArgs):
     sglang_args = (
         "--rollout-num-gpus-per-engine 8 "
         f"--sglang-mem-fraction-static 0.7 {sglang_extra}"
-        "--sglang-cuda-graph-max-bs 512 "
+        "--sglang-cuda-graph-max-bs-decode 512 "
     )
 
     misc_args = (

diff --git a/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh b/examples/infra_features/low_precision/run-qwen3-235B-A22B-int4.sh
@@ -118,7 +118,7 @@ SGLANG_ARGS=(
   #  --sglang-enable-dp-attention
   #  --sglang-dp-size 4
    --sglang-ep-size 8
-   --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
+   --sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)
 )
 
 

diff --git a/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh b/examples/infra_features/low_precision/run-qwen3-30B-A3B-int4.sh
@@ -113,7 +113,7 @@ WANDB_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 1
    --sglang-mem-fraction-static 0.7
-   --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
+   --sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)
 )
 
 MISC_ARGS=(

diff --git a/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh b/examples/infra_features/low_precision/run-qwen3-30b-a3b-fp8-two-nodes.sh
@@ -126,7 +126,7 @@ WANDB_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8
    --sglang-mem-fraction-static 0.6
-   --sglang-cuda-graph-bs 1 2 4 8 $(seq 16 8 256)
+   --sglang-cuda-graph-bs-decode 1 2 4 8 $(seq 16 8 256)
    --sglang-expert-parallel-size 8
    # --use-rollout-routing-replay
 )

diff --git a/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm4.5-air-8node-profile.sh
@@ -179,7 +179,7 @@ run_mode() {
         --rollout-num-gpus ${NUM_ROLLOUT_GPUS}
         --sglang-mem-fraction-static 0.8
         --sglang-ep-size 8
-        --sglang-cuda-graph-bs 1 2 4 8 16
+        --sglang-cuda-graph-bs-decode 1 2 4 8 16
         --sglang-enable-dp-attention
         --sglang-enable-dp-lm-head
     )

diff --git a/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm4.7-flash-2node-profile.sh
@@ -180,7 +180,7 @@ run_mode() {
         --rollout-num-gpus ${NUM_ROLLOUT_GPUS}
         --sglang-mem-fraction-static 0.7
         --sglang-ep-size 4
-        --sglang-cuda-graph-bs 1 2 4 8 16
+        --sglang-cuda-graph-bs-decode 1 2 4 8 16
         --sglang-enable-dp-attention
         --sglang-enable-dp-lm-head
     )

diff --git a/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh b/examples/infra_features/p2p_weight_transfer/run-glm5-disagg-profile.sh
@@ -221,10 +221,10 @@ run_mode() {
         --sglang-enable-dp-lm-head
         # GLM5 NSA attention and chunked prefill
         --sglang-page-size 64
-        --sglang-nsa-decode-backend flashmla_sparse
-        --sglang-nsa-prefill-backend flashmla_sparse
+        --sglang-dsa-decode-backend flashmla_sparse
+        --sglang-dsa-prefill-backend flashmla_sparse
         --sglang-attention-backend nsa
-        --sglang-cuda-graph-max-bs ${SGLANG_DECODE_MAX_BS}
+        --sglang-cuda-graph-max-bs-decode ${SGLANG_DECODE_MAX_BS}
         --sglang-max-running-requests 512
         --sglang-chunked-prefill-size $((2048 * ${SGLANG_WORLD_SIZE}))
         --sglang-watchdog-timeout 3600

diff --git a/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-kimi-k2-64node-profile.sh
@@ -190,7 +190,7 @@ run_mode() {
         --sglang-enable-dp-attention
         --sglang-enable-dp-lm-head
         --sglang-server-concurrency 1024
-        --sglang-cuda-graph-bs 1 2 4 8 16
+        --sglang-cuda-graph-bs-decode 1 2 4 8 16
         --sglang-moe-runner-backend triton
         --sglang-fp8-gemm-backend triton
     )

diff --git a/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-qwen3-235B-A22B-16node-profile.sh
@@ -181,7 +181,7 @@ run_mode() {
         --sglang-mem-fraction-static 0.75
         --sglang-ep-size 32
         --sglang-dp-size 1
-        --sglang-cuda-graph-bs 1 2 4 8 16
+        --sglang-cuda-graph-bs-decode 1 2 4 8 16
         --sglang-enable-dp-attention
         --sglang-enable-dp-lm-head
     )

diff --git a/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh b/examples/infra_features/p2p_weight_transfer/run-qwen3-30B-A3B-4node-profile.sh
@@ -185,7 +185,7 @@ run_mode() {
         --rollout-num-gpus ${NUM_ROLLOUT_GPUS}
         --sglang-mem-fraction-static 0.8
         --sglang-ep-size 8
-        --sglang-cuda-graph-bs 1 2 4 8 16
+        --sglang-cuda-graph-bs-decode 1 2 4 8 16
         --sglang-enable-dp-attention
         --sglang-enable-dp-lm-head
     )

diff --git a/examples/infra_features/p2p_weight_transfer/run.py b/examples/infra_features/p2p_weight_transfer/run.py
@@ -937,7 +937,7 @@ def cmd_run(
     if cfg.sglang_moe_dense_tp_size is not None:
         args.extend(["--sglang-moe-dense-tp-size", str(cfg.sglang_moe_dense_tp_size)])
     if cfg.sglang_cuda_graph_bs:
-        args.extend(["--sglang-cuda-graph-bs"] + cfg.sglang_cuda_graph_bs.split())
+        args.extend(["--sglang-cuda-graph-bs-decode"] + cfg.sglang_cuda_graph_bs.split())
     if cfg.sglang_enable_dp_attention:
         args.append("--sglang-enable-dp-attention")
     if cfg.sglang_enable_dp_lm_head:
@@ -955,16 +955,16 @@ def cmd_run(
             args.extend(["--sglang-page-size", str(cfg.sglang_page_size)])
         args.extend(
             [
-                "--sglang-nsa-decode-backend",
+                "--sglang-dsa-decode-backend",
                 "flashmla_sparse",
-                "--sglang-nsa-prefill-backend",
+                "--sglang-dsa-prefill-backend",
                 "flashmla_sparse",
                 "--sglang-attention-backend",
                 "nsa",
             ]
         )
         if cfg.sglang_cuda_graph_max_bs is not None:
-            args.extend(["--sglang-cuda-graph-max-bs", str(cfg.sglang_cuda_graph_max_bs)])
+            args.extend(["--sglang-cuda-graph-max-bs-decode", str(cfg.sglang_cuda_graph_max_bs)])
         if cfg.sglang_max_running_requests is not None:
             args.extend(["--sglang-max-running-requests", str(cfg.sglang_max_running_requests)])
         if cfg.sglang_chunked_prefill_size_factor is not None:

diff --git a/examples/lora/run-kimi-k25-megatron-lora.sh b/examples/lora/run-kimi-k25-megatron-lora.sh
@@ -134,7 +134,7 @@ SGLANG_ARGS=(
    --sglang-mem-fraction-static 0.7
    --sglang-ep-size 8
    --sglang-server-concurrency 1024
-   --sglang-cuda-graph-bs 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 24 32 40 48 56 64 72 80 88 96 104 112 120 128
    --use-rollout-routing-replay
 )
 

diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase1_rlvr_teacher.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase1_rlvr_teacher.sh
@@ -83,8 +83,8 @@ PERF_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.8 --sglang-ep-size 8
    --sglang-watchdog-timeout 1800 --sglang-enable-metrics --sglang-attention-backend fa3
-   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
-   --sglang-mamba-scheduler-strategy extra_buffer
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-radix-cache-strategy extra_buffer
 )
 MISC_ARGS=(
    --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32

diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_gb200.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_gb200.sh
@@ -109,8 +109,8 @@ SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.7 --sglang-ep-size 8
    --sglang-watchdog-timeout 1800 --sglang-enable-metrics
    --sglang-moe-runner-backend flashinfer_cutlass --sglang-attention-backend trtllm_mha
-   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
-   --sglang-mamba-scheduler-strategy extra_buffer
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-radix-cache-strategy extra_buffer
 )
 MISC_ARGS=(
    --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32

diff --git a/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_opd_selfdistill.sh b/examples/on_policy_distillation/qwen3_5_35b_selfdistill/phase2_opd_selfdistill.sh
@@ -98,8 +98,8 @@ PERF_ARGS=(
 SGLANG_ARGS=(
    --rollout-num-gpus-per-engine 8 --sglang-mem-fraction-static 0.8 --sglang-ep-size 8
    --sglang-watchdog-timeout 1800 --sglang-enable-metrics --sglang-attention-backend fa3
-   --sglang-cuda-graph-bs 1 2 4 8 16 32 --use-rollout-routing-replay
-   --sglang-mamba-scheduler-strategy extra_buffer
+   --sglang-cuda-graph-bs-decode 1 2 4 8 16 32 --use-rollout-routing-replay
+   --sglang-mamba-radix-cache-strategy extra_buffer
 )
 MISC_ARGS=(
    --attention-dropout 0.0 --hidden-dropout 0.0 --accumulate-allreduce-grads-in-fp32