feat(examples/strands_sglang): update to strands-sglang 0.4.2 (#2106)

Co-authored-by: YuanHe <lawhy@amazon.com>
Co-authored-by: Chengxing Xie <xiechengxing34@gmail.com>

Author: 3 people

examples/strands_sglang/README.md

@@ -13,7 +13,7 @@ This example connects `slime` with [`strands-sglang`](https://github.com/horizon
 `strands-sglang` bridges the gap by extending `strands` with SGLang's native `/generate` endpoint:
 
 - Captures exact token IDs during generation (no retokenization drift)
-- Automatically tracks `loss_mask` via `token_manager`
+- Automatically tracks `loss_mask` via the `Rollout` tracker (`model.rollout`)
 - Provides `ToolLimiter` for clean trajectory truncation
 
 ## Install Dependencies
@@ -22,11 +22,9 @@ This example connects `slime` with [`strands-sglang`](https://github.com/horizon
 2. Go to slime folder: `cd /root/slime`
 3. Install slime: `pip install -e . --no-deps`
 4. Go to the example folder: `cd /root/slime/examples/strands_sglang`
-5. Install other dependencies: `pip install -r requirements.txt`
+5. Install `strands-sglang`: `pip install strands-sglang==0.4.2`
 
-> NOTE: `strands-sglang` is under rapid development, so we recommend using the GitHub repo version: `strands-sglang @ git+https://github.com/horizon-rl/strands-sglang.git`
-
-> NOTE: We use camel-ai's subprocess code interpreter for python code execution, which is NOT a good practice; it's just for convenience of this example.
+> NOTE: The `execute_python_code` tool runs code via `subprocess_interpreter.py`, a self-contained interpreter vendored from camel-ai so this example does not depend on the full `camel-ai` package. It runs model-generated code in a local subprocess with **no isolation**, which is NOT a good practice; it is here only for the convenience of this example. Use a sandboxed interpreter (Docker, e2b, microsandbox, ...) for anything beyond local experimentation.
 
 ## Prepare Model
 

examples/strands_sglang/__init__.py

@@ -1,7 +1,5 @@
-# Updated with strands-sglang 0.3.2
 import logging
 
-from camel.interpreters import SubprocessInterpreter
 from strands import Agent, tool
 from strands_sglang import SGLangModel, ToolLimiter, get_client_from_slime_args
 from strands_sglang.tool_parsers import HermesToolParser
@@ -10,6 +8,8 @@
 from slime.rollout.sglang_rollout import GenerateState
 from slime.utils.types import Sample
 
+from .subprocess_interpreter import SubprocessInterpreter
+
 logger = logging.getLogger(__name__)
 
 SYSTEM_PROMPT = """
@@ -22,7 +22,6 @@
 """.strip()
 
 MAX_TOOL_ITERS = 5
-MAX_TOOL_CALLS = None  # No limit
 
 
 @tool
@@ -51,7 +50,7 @@ async def generate(args, sample: Sample, sampling_params) -> Sample:
         sampling_params=sampling_params,
     )
 
-    tool_limiter = ToolLimiter(max_tool_iters=MAX_TOOL_ITERS, max_tool_calls=MAX_TOOL_CALLS)
+    tool_limiter = ToolLimiter(max_tool_iters=MAX_TOOL_ITERS)
     agent = Agent(
         model=model,
         tools=[execute_python_code],
@@ -71,12 +70,12 @@ async def generate(args, sample: Sample, sampling_params) -> Sample:
         sample.status = Sample.Status.TRUNCATED
         logger.warning(f"TRUNCATED: {type(e).__name__}: {e}")
 
-    # Extract token trajectory from token_manager
-    tm = model.token_manager
-    prompt_len = len(tm.segments[0])  # system + user are first segment
-    sample.tokens = tm.token_ids
-    sample.loss_mask = tm.loss_mask[prompt_len:]
-    sample.rollout_log_probs = tm.logprobs[prompt_len:]
+    # Extract token trajectory from the rollout tracker
+    rollout = model.rollout
+    prompt_len = rollout.initial_prompt_length  # system + user are the first segment
+    sample.tokens = rollout.token_ids
+    sample.loss_mask = rollout.loss_mask[prompt_len:]
+    sample.rollout_log_probs = rollout.logprobs[prompt_len:]
     sample.response_length = len(sample.tokens) - prompt_len
     sample.response = model.tokenizer.decode(sample.tokens[prompt_len:], skip_special_tokens=False)
     # Tool iteration and tool call count are different because multiple parallel tool calls count as 1 iteration