[TRTLLM-13233][feat] Support no_repeat_ngram_size in TorchSampler

docs/source/features/sampling.md

@@ -115,6 +115,10 @@ llm.generate(["Hello, my name is",
   * Top-P decay is not supported in combination with beam search or with speculative decoding
     modes that route draft tokens through the Torch Sampler; such requests are rejected.
 
+* If `no_repeat_ngram_size = n` is specified, any token that would recreate an `n`-gram already
+  present in the sequence (prompt included) is excluded from sampling. `None` or `0` disables
+  the restriction.
+
 ### Performance
 
 The Torch Sampler leverages the optimized sampling kernels provided by

tensorrt_llm/_torch/pyexecutor/llm_request.py

@@ -1201,6 +1201,13 @@ def executor_request_to_llm_request(
         list(word) for word in executor_request.bad_words
     ] if executor_request.bad_words else None
 
+    # No-repeat-ngram size for the TorchSampler path, normalized once here so
+    # the per-step sampler gate is a plain attribute read. The C++
+    # SamplingConfig rejects negative values up front and 0 means disabled
+    # (same convention as the C++ banRepeatNgram kernel), so falsy == off.
+    ngram_size = executor_request.sampling_config.no_repeat_ngram_size
+    llm_request.py_no_repeat_ngram_size = ngram_size if ngram_size else None
+
     llm_request.py_original_end_id = getattr(executor_request,
                                              "py_original_end_id",
                                              llm_request.py_end_id)