feature: make trtllmsampler new_tokens format the universal format

[netanel-haber]

No description provided.

[netanel-haber]

/bot run

[tensorrt-cicd]

PR_Github #6632 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #6632 [ run ] completed with state FAILURE
/LLM/main/L0_MergeRequest_PR pipeline #4848 completed with status: 'FAILURE'

[netanel-haber]

/bot run

[tensorrt-cicd]

PR_Github #7516 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7516 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5451 completed with status: 'FAILURE'

[netanel-haber]

/bot run

[tensorrt-cicd]

PR_Github #7698 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7698 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5588 completed with status: 'FAILURE'

[netanel-haber]

/bot run

[tensorrt-cicd]

PR_Github #9539 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #9539 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #7000 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.

[Copilot]

Pull Request Overview

This PR refactors how speculative samplers handle new token formatting by unifying on a single TorchSampler.Args structure, streamlining decoder factory logic, and replacing legacy sampler implementations.

• Refactored get_spec_decoder to accept TorchSampler.Args and updated MTP/Eagle3OneModel sampler constructors.
• Consolidated request iteration via ScheduledRequests.all_requests(), replacing itertools.chain across the codebase.
• Removed outdated Eagle3Sampler/Eagle3Decoder classes and integrated SeqSlotManager for draft slot management.

Reviewed Changes

Copilot reviewed 11 out of 12 changed files in this pull request and generated no comments.

Show a summary per file

File	Description
speculative/utils.py	Updated decoder factory signature and imports for spec decoders
speculative/mtp.py	Refactored MTPSampler constructor, updated stop criteria logic
speculative/eagle3.py	Removed legacy sampler/decoder classes, added new sampler class
pyexecutor/seq_slot_manager.py	Simplified resource prep using all_requests()
pyexecutor/scheduler.py	Changed all_requests to a method returning a list
pyexecutor/py_executor.py	Integrated SeqSlotManager, updated logits field assignments
pyexecutor/model_engine.py	Introduced BEAM_WIDTH, centralized batch indexing logic
pyexecutor/llm_request.py	Added py_is_draft flag to LlmRequest
pyexecutor/guided_decoder.py	Replaced itertools.chain with all_requests()
pyexecutor/_util.py	Centralized sampler instantiation with create_torch_sampler_args
auto_deploy/shim/ad_executor.py	Updated AD executor to use TorchSampler.Args and slot manager

Comments suppressed due to low confidence (3)

tensorrt_llm/_torch/speculative/mtp.py:314

• The returned SampleStateMTP no longer includes a logits field, which may be accessed downstream in the executor (e.g., in _executor_loop_pp). Consider preserving or setting device.logits and host.logits in SampleStateMTP to avoid missing attribute errors.

        )

tensorrt_llm/_torch/speculative/utils.py:83

• [nitpick] The parameter name sampler_args is more verbose than other code that uses args for TorchSampler parameters. Consider renaming it to args for consistency and brevity.

def get_spec_decoder(sampler_args: TorchSampler.Args, spec_config: SpecConfig):

tensorrt_llm/_torch/pyexecutor/model_engine.py:1160

• [nitpick] The nonlocal mtp_batch_idx declaration appears after a conditional return in the nested py_batch_idx function. For clarity, move the nonlocal statement to the top of the function body before any logic.

            nonlocal mtp_batch_idx

[suyoggupta]

AD changes LGTM