From b9599194042ebbf1c73daad71858fa06a7f962d1 Mon Sep 17 00:00:00 2001
From: trtllm-agent <296075020+trtllm-agent@users.noreply.github.com>
Date: Mon, 31 Aug 2026 23:13:23 -0700
Subject: [PATCH] [https://nvbugs/6693991][fix] Tolerate the removed
 sampler_type LLM arg

Removing sampler_type left no back-compat path, so callers pinned to an
older TRT-LLM -- notably the verl rollout server, which passes it
unconditionally on every version including HEAD -- died in
_validate_args_for_torch_backend before the engine was ever built.

Absorb the key with a warning instead. TorchSampler is now the only
sampler, so every former value resolves to the same behavior and there is
no choice left to honor; raising would only relocate the same hard
failure. Both pre-pydantic kwarg gates consult the removed-key set, since
they reject unknown names before the validator can run.

No field is reintroduced, so the api-stability reference and the golden
telemetry manifest are unchanged.

Signed-off-by: trtllm-agent <296075020+trtllm-agent@users.noreply.github.com>
---
 tensorrt_llm/llmapi/llm.py      | 11 +++++++----
 tensorrt_llm/llmapi/llm_args.py | 28 ++++++++++++++++++++++++++++
 2 files changed, 35 insertions(+), 4 deletions(-)

diff --git a/tensorrt_llm/llmapi/llm.py b/tensorrt_llm/llmapi/llm.py
index db866443236e..53f4e40504a6 100644
--- a/tensorrt_llm/llmapi/llm.py
+++ b/tensorrt_llm/llmapi/llm.py
@@ -57,7 +57,8 @@
 from ..logger import logger
 from ..sampling_params import LogitsProcessor, SamplingParams
 from ..scheduling_params import SchedulingParams
-from .llm_args import (TORCH_LLMARGS_EXPLICIT_DOCSTRING, TorchLlmArgs,
+from .llm_args import (TORCH_LLMARGS_EXPLICIT_DOCSTRING,
+                       TORCH_LLMARGS_REMOVED_KEYS, TorchLlmArgs,
                        validate_token_encoder_bucket_config)
 from .llm_utils import (CachedModelLoader, KvCacheRetentionConfig,
                         LlmBuildStats, ModelLoader)
@@ -403,6 +404,8 @@ def __init__(self,
             valid_keys = set(
                 list(llm_args_cls.model_fields.keys()) +
                 ['_mpi_session', 'backend'])
+            if issubclass(llm_args_cls, TorchLlmArgs):
+                valid_keys |= TORCH_LLMARGS_REMOVED_KEYS
             for key in kwargs:
                 if key not in valid_keys:
                     raise ValueError(
@@ -2012,9 +2015,9 @@ def _validate_args_for_torch_backend(self, kwargs: dict) -> None:
 
         # Check if any arguments not supported by the PyTorch backend are passed.
         unsupported_args = [
-            key for key in kwargs
-            if key not in torchllm_fields and key not in ('_mpi_session',
-                                                          'backend')
+            key for key in kwargs if key not in torchllm_fields and key not in (
+                '_mpi_session',
+                'backend') and key not in TORCH_LLMARGS_REMOVED_KEYS
         ]
 
         if unsupported_args:
diff --git a/tensorrt_llm/llmapi/llm_args.py b/tensorrt_llm/llmapi/llm_args.py
index 1ffb613eb660..7feec022b9c0 100644
--- a/tensorrt_llm/llmapi/llm_args.py
+++ b/tensorrt_llm/llmapi/llm_args.py
@@ -5294,7 +5294,35 @@ def validate_capture_num_tokens(cls, v):
         "The maximum number of CUDA streams to use for torch.compile.")
 
 
+# Keys that were removed from TorchLlmArgs but are still tolerated on input, so
+# that callers pinned to an older TRT-LLM keep constructing. `_TorchLLM` and
+# `BaseLLM` reject unknown kwargs before pydantic runs, so both consult this set
+# too. Each entry is absorbed by a `mode="before"` validator below.
+TORCH_LLMARGS_REMOVED_KEYS = frozenset({"sampler_type"})
+
+
 class TorchLlmArgs(BaseLlmArgs):
+
+    @model_validator(mode="before")
+    @classmethod
+    def _warn_removed_sampler_type(cls, data):
+        """Absorb the removed `sampler_type` knob.
+
+        TorchSampler is now the only sampler, so sampler selection has no
+        meaning and every former value resolves to the same behavior. Drop the
+        key with a warning rather than raising: raising would break the pinned
+        RL/rollout integrations (e.g. verl) that still pass it, for a choice
+        they can no longer make either way.
+        """
+        if isinstance(data, dict) and "sampler_type" in data:
+            data = dict(data)
+            requested = data.pop("sampler_type")
+            logger.warning(
+                f"'sampler_type' was removed (got {requested!r}) and is "
+                "ignored. TorchSampler is the only sampler; remove the "
+                "argument to silence this warning.")
+        return data
+
     # PyTorch backend specific configurations
     generation_config: Literal["auto", "trtllm"] = Field(
         default="trtllm",