awslabs · lliquid · Jul 3, 2026 · Jul 2, 2026 · Jul 2, 2026
diff --git a/docs/site/scripts/gen_api.py b/docs/site/scripts/gen_api.py
@@ -20,12 +20,7 @@
         reward.md
       backends/
         slime/
-          rollout.md
-          traces.md
-          rewards.md
-          gateway.md
-          sglang_token_ids.md
-          verify_sglang_token_ids.md
+          runner.md
 """
 from __future__ import annotations
 
@@ -98,32 +93,17 @@ class ModuleSpec:
         include=("RewardFunction",),
     ),
     # --- Backends › slime --- #
-    # Only the public user surface: the SlimeRunner entry point and the
-    # two SGLang patch scripts. integration/rollout.py and
-    # integration/rewards.py are load-bearing plugin paths but users
-    # don't import them — SlimeRunner wires them into the job.
+    # Only the public user surface: the SlimeRunner entry point.
+    # integration/rollout.py and integration/rewards.py are load-bearing
+    # plugin paths but users don't import them — SlimeRunner wires them
+    # into the job.
     ModuleSpec(
         dotted_paths=("agentcore_rl_toolkit.backends.slime.runner",),
         out_path="backends/slime/runner.md",
         title="SlimeRunner",
         description="One Python entry point for slime-backed training.",
         include=("SlimeRunner",),
     ),
-    # Both patches consolidated onto one page.
-    # sglang_token_ids exposes apply_patch (the public patch entry point).
-    # verify_sglang_token_ids exposes main (the CLI contract, invoked via
-    # `python -m ...verify_sglang_token_ids`). Everything else is
-    # low-level monkey-patch machinery / internal check helpers.
-    ModuleSpec(
-        dotted_paths=(
-            "agentcore_rl_toolkit.backends.slime.patches.sglang_token_ids",
-            "agentcore_rl_toolkit.backends.slime.patches.verify_sglang_token_ids",
-        ),
-        out_path="backends/slime/patches.md",
-        title="SGLang patches",
-        description="Apply + verify the SGLang token-IDs patch used by the slime backend.",
-        include=("apply_patch", "main"),
-    ),
 ]
 
 

diff --git a/docs/site/src/content/docs/api/backends/slime/patches.md b/docs/site/src/content/docs/api/backends/slime/patches.md
diff --git a/docs/site/src/content/docs/api/backends/slime/runner.md b/docs/site/src/content/docs/api/backends/slime/runner.md
@@ -41,13 +41,22 @@ SlimeRunner(
     slime_dir: str = '/root/slime',
     megatron_dir: str = '/root/Megatron-LM',
 
+    # --- Optional: CUDA toolchain pinning ---
+    cuda_home: str | None = None,
+
     # --- Optional: ACR / toolkit (forwarded to slime via custom-config yaml) ---
     model_id: str = 'default',
     acr_timeout: int = 900,
     acr_tps_limit: int = 25,
     max_concurrent: int = 100,
+    max_pool_connections: int = 100,
     gateway_port: int = 9090,
     reward_postprocessing: str = 'grpo',
+    sglang_tool_call_parser: str = 'qwen',
+    sglang_reasoning_parser: str | None = None,
+    cumulative_token_mode: bool = False,
+    renderer_family: str = 'auto',
+    gateway_log_level: str = 'warning',
 
     # --- Optional: training hyperparameters ---
     rollout_batch_size: int = 32,
@@ -60,6 +69,7 @@ SlimeRunner(
     weight_decay: float = 0.1,
     adam_beta2: float = 0.98,
     sglang_mem_fraction_static: float = 0.7,
+    sglang_context_length: int | None = None,
     max_tokens_per_gpu: int = 9216,
 
     # --- Wandb (opt-in; no defaults injected if unset) ---
@@ -95,6 +105,10 @@ via ``ray job submit``. Streams stdout/stderr to the parent process.
 
 - `agent_runtime_arn` *(str)*
 
+- `cuda_home` *(str | None)*
+
+- `cumulative_token_mode` *(bool)*
+
 - `data_path` *(str)*
 
 - `eps_clip` *(float)*
@@ -105,12 +119,16 @@ via ``ray job submit``. Streams stdout/stderr to the parent process.
 
 - `extra_flags` *(list[str])*
 
+- `gateway_log_level` *(str)*
+
 - `gateway_port` *(int)*
 
 - `lr` *(float)*
 
 - `max_concurrent` *(int)*
 
+- `max_pool_connections` *(int)*
+
 - `max_tokens_per_gpu` *(int)*
 
 - `megatron_dir` *(str)*
@@ -125,6 +143,8 @@ via ``ray job submit``. Streams stdout/stderr to the parent process.
 
 - `num_gpus` *(int)*
 
+- `renderer_family` *(str)*
+
 - `reward_postprocessing` *(str)*
 
 - `rollout_batch_size` *(int)*
@@ -137,8 +157,14 @@ via ``ray job submit``. Streams stdout/stderr to the parent process.
 
 - `s3_bucket` *(str)*
 
+- `sglang_context_length` *(int | None)*
+
 - `sglang_mem_fraction_static` *(float)*
 
+- `sglang_reasoning_parser` *(str | None)*
+
+- `sglang_tool_call_parser` *(str)*
+
 - `slime_dir` *(str)*
 
 - `tp_size` *(int)*

diff --git a/docs/site/src/content/docs/guides/slime-backend-setup.md b/docs/site/src/content/docs/guides/slime-backend-setup.md
@@ -5,23 +5,18 @@ description: Train an AgentCore Runtime-deployed agent with the SLIME training b
 
 This doc describes how to train an AgentCore Runtime-deployed agent with the
 [slime](https://github.com/THUDM/slime) training backend. The public
-user surface is exactly two things: the
+user surface is the
 [`SlimeRunner`](/agentcore-rl-toolkit/api/backends/slime/runner/)
-class (for launching training) and the two SGLang
-[patch scripts](/agentcore-rl-toolkit/api/backends/slime/patches/)
-(applied once to the SGLang install).
+class for launching training.
 
-For known issues (Megatron-LM regression on 32B, norm-epsilon
-mismatch, etc.) see
+For known issues (e.g. the norm-epsilon mismatch on
+Qwen2.5-32B-Instruct) see
 [slime troubleshooting](/agentcore-rl-toolkit/troubleshooting/slime/).
 
 ## Prerequisites
 
-- Hardware and CUDA requirements: see
-  [slime's README](https://github.com/THUDM/slime#installation) and the
-  [slime docker README](https://github.com/THUDM/slime/blob/main/docker/README.md)
-  for tested GPU configurations per model size.
-- Python 3.10+ and [`uv`](https://docs.astral.sh/uv/).
+- A GPU cluster with **CUDA>=12.9** installed.
+- Python 3.12+ and [`uv`](https://docs.astral.sh/uv/).
 - AWS credentials with permission to invoke an AgentCore Runtime and
   read/write an S3 bucket.
 - An AgentCore Runtime deployment of your agent — follow the
@@ -31,44 +26,52 @@ mismatch, etc.) see
 - An S3 bucket for rollout result delivery — required as the
   `s3_bucket` argument on `SlimeRunner` below.
 
-## slime environment
+## Installation
 
-Follow
-[slime's own installation docs](https://github.com/THUDM/slime#installation)
-— either the container path (`slimerl/slime:latest`) or a bare-metal
-install. Everything below runs inside this environment.
+Choose **one** of the two paths below to install slime, then install
+the toolkit into the same environment.
 
-Inside the `slimerl/slime:latest` container, slime and Megatron-LM
-ship pre-installed at `/root/slime` and `/root/Megatron-LM` — use
-those paths for `slime_dir` / `megatron_dir` on `SlimeRunner`. For a
-bare-metal install, point at wherever you cloned slime + Megatron-LM.
+### Option A: Official slime docker
 
-## Install the toolkit + apply patches
+Follow
+[slime's own installation docs](https://github.com/THUDM/slime#installation)
+and use the container image (`slimerl/slime:latest`). Inside the
+container, slime and Megatron-LM ship pre-installed at `/root/slime`
+and `/root/Megatron-LM` — use those paths for `slime_dir` /
+`megatron_dir` on `SlimeRunner`.
 
-Inside the slime environment:
+Install the toolkit with the slime-backend extras inside the container:
 
 ```bash
-# From a clone of this repo
-cd /path/to/agentcore-rl-toolkit
-
-# Install the toolkit plus the slime-backend extras
 uv pip install -e ".[slime]"
 ```
 
-Then apply the SGLang `token_ids` patch — it adds
-`prompt_token_ids` / `token_ids` fields to chat completion responses
-so the gateway can capture RL training trace data. The patch is idempotent:
+:::note
+We have only tested against official slime at commit
+[`fa3c990`](https://github.com/THUDM/slime/commit/fa3c990af6f18efd3fd9922698bf4bf4048d1263).
+:::
 
-```bash
-python -m agentcore_rl_toolkit.backends.slime.patches.sglang_token_ids
+### Option B: Bare-metal install script
 
-# Verify the patch round-trips under greedy decoding (any HF checkpoint
-# works; Qwen2.5-0.5B-Instruct is the fastest to download + load)
-python -m agentcore_rl_toolkit.backends.slime.patches.verify_sglang_token_ids \
-    --model-path /path/to/Qwen2.5-0.5B-Instruct
-# Expect: "OK: 4/4 checks passed"
+Install slime and its heavyweight dependency stack (Megatron-LM,
+Transformer Engine, Apex, flash-attn, sglang, torch_memory_saver)
+with the provided script, which clones slime + Megatron-LM into the
+current directory and applies slime's official patches. Run it inside
+your activated python environment.
+
+:::note
+Both CUDA 12.9 and CUDA 13 are supported. The following commands assume CUDA 13.0 is installed at `/usr/local/cuda-13.0`; adjust `CUDA_HOME` and `cu13` if yours differs.
+:::
+
+```bash
+uv pip install -e ".[slime]"
+export CUDA_HOME=/usr/local/cuda-13.0
+bash src/agentcore_rl_toolkit/backends/slime/scripts/install_slime.sh cu13
 ```
 
+Point `slime_dir` / `megatron_dir` on `SlimeRunner` at the `slime` and
+`Megatron-LM` directories the script cloned.
+
 ## Prepare data
 
 The training dataset is a JSONL file where each line is one rollout

diff --git a/docs/site/src/content/docs/troubleshooting/slime.md b/docs/site/src/content/docs/troubleshooting/slime.md
@@ -14,51 +14,11 @@ was validated against:
 | Component | Version / SHA |
 |---|---|
 | Instance type | 8 × NVIDIA H100 80GB HBM3 |
-| CUDA | `12.9` |
-| PyTorch | `2.9.1+cu129` |
-| Docker image | `slimerl/slime@sha256:0100c933f1f63e7c4acdb9ec575e769839d59de4a648551e09e3fe0e7885631b` (built 2026-04-28) |
-| slime | commit `f3e7bd7f3091d3be05c20977eefb31a785d6221d` (2026-04-28) |
-| SGLang | `v0.5.9` |
-| Megatron-LM | commit `3714d81d418c9f1bca4594fc35f9e8289f652862` ⚠ see note |
-
-:::caution[Megatron-LM pin]
-The image bundles Megatron-LM at `1dcf0dafa` (~500 commits ahead of
-slime's stable pin), which breaks 32B training — see the
-[LinearCrossEntropyModule issue](#linearcrossentropymodule-parallelism-error-on-32b-or-any-model-with-untied-embeddings) below.
-We downgrade to `3714d81d` (slime's documented stable sha) via
-`git checkout` inside `/root/Megatron-LM`. The table above reflects
-the downgraded sha, not the one baked into the image.
-:::
-
-## `LinearCrossEntropyModule` parallelism error on 32B (or any model with untied embeddings)
-
-**Symptom:** During 32B training, the Megatron actor crashes with:
-
-```
-ValueError: Cannot determine parallelism type for module 'LinearCrossEntropyModule'
-            at weight 'output_layer.weight'.
-```
-
-**Cause:** The Megatron-LM bundled in `slimerl/slime:latest` is
-several hundred commits ahead of the sha pinned in slime's docker
-README (`3714d81d`). Specifically, Megatron PR **#3226 "Reapply fix
-Linear CE Fusion"** (2026-02-04) replaced `ColumnParallelLinear`
-with a new `LinearCrossEntropyModule` that megatron-bridge's
-`AutoMapping` doesn't recognize. Models with tied embeddings (0.5B,
-3B, 7B) skip this code path; models with
-`--untie-embeddings-and-output-weights` (32B and up) hit it.
-
-**Fix:** Inside the container, pin `/root/Megatron-LM` to the stable
-sha:
-
-```bash
-cd /root/Megatron-LM
-# Stash any image-local patches first (can be restored later with `git stash pop`)
-git stash -u -m "slime local patches"
-git checkout 3714d81d418c9f1bca4594fc35f9e8289f652862
-# Clear pyc caches that reference the old code
-find . -name __pycache__ -type d -exec rm -rf {} + 2>/dev/null
-```
+| CUDA | `13.0` |
+| PyTorch | `2.11.0+cu130` |
+| slime | commit `fa3c990af6f18efd3fd9922698bf4bf4048d1263` |
+| SGLang | `0.5.13` |
+| Megatron-LM | commit `1dcf0dafa884ad52ffb243625717a3471643e087` |
 
 ## `--norm-epsilon` mismatch on Qwen2.5-32B-Instruct
 
@@ -68,5 +28,5 @@ find . -name __pycache__ -type d -exec rm -rf {} + 2>/dev/null
 
 **Fix:** Edit the slime model script to `--norm-epsilon 1e-6`, or
 pass an override through `SlimeRunner(extra_flags=["--norm-epsilon", "1e-6"])`.
-0.5B/3B/7B Instruct variants match their base-model norm epsilons,
-so this only affects 32B.
+Qwen2.5-0.5B / 1.5B / 3B / 7B Instruct variants match their base-model
+norm epsilons, so this only affects Qwen2.5-32B.