Repository navigation

[TRTLLM-6785][feat] BREAKING CHANGE Enable TRTLLM sampler by default - #6216

Merged

dcampora merged 32 commits into

NVIDIA:mainfrom

dcampora:user/dcampora/enable_trtllm_sampler_default

Aug 8, 2025

Merged

[TRTLLM-6785][feat] BREAKING CHANGE Enable TRTLLM sampler by default#6216
dcampora merged 32 commits into
NVIDIA:mainfrom
dcampora:user/dcampora/enable_trtllm_sampler_default

Commits on Aug 7, 2025
- Make TRTLLM Sampler the default.
  dcampora
  committed
  a49bcf4
- Fix default.
  dcampora
  committed
  69b2c7f
- Fix test_return_logits.
  dcampora
  committed
  769051b
- Wider support of executor_config.speculative_config.
  dcampora
  committed
  d47ede0
- Set max batch size to 64 in test_llm_api_pytorch qwen
  dcampora
  committed
  52e2346
- Fix run of tests against spec decoding)
  dcampora
  committed
  caf4b77
- Fix max in max_attn_window values.
  dcampora
  committed
  65c45d6
- Formatting.
  dcampora
  committed
  208c67c
- Remove enable_trtllm_sampler option.
  dcampora
  committed
  bf152cf
- Unwaive qwen test, adapt sampler output.
  dcampora
  committed
  7519e05
- Adapt e2e test format.
  dcampora
  committed
  560ca3a
- Use Qwen3/Qwen3-0.6B-Base instead of TinyLlama for its longer max_position_embeddings.
  dcampora
  committed
  4c7c457
- Remove unnecessary enable_torch_sampler in test.
  dcampora
  committed
  6d37840
- Removed unnecessary setting of Torch sampler to false.
  dcampora
  committed
  e115b5a
- Change enable_torch_sampler to use_torch_sampler.
  dcampora
  committed
  4549ee5
- Fix remaining single gpu issues.
  dcampora
  committed
  78f065a
- Fix remaining failing tests.
  dcampora
  committed
  a8cea27
- Apply test correction.
  dcampora
  committed
  ae1a128
- test_disaggregated_torch_sampler
  dcampora
  committed
  e804d4d
- Use torch sampler for tests that use large models.
  dcampora
  committed
  9e41f73
- refactor: Rename maxBatchSize to maxNumSequences in decoder and gptDecoder classes
  
  Funatiq
  authored and
  dcampora
  committed
  f03203a
- finalize_events: dict[str, CudaEvent]** | None = None**
  
  netanel-haber
  authored and
  dcampora
  committed
  810d0ff
- wording
  
  netanel-haber
  authored and
  dcampora
  committed
  435d19e
- Adapt to use_torch_sampler.
  dcampora
  committed
  5a5d642
- Fix test_openai_chat_multimodal.
  dcampora
  committed
  2e653a1
- setup is_attention_dp_dummy=True gen requests
  
  netanel-haber
  authored and
  dcampora
  committed
  41ff222
- Use torch sampler in multimodal.
  dcampora
  committed
  cd1c497
- minimal disagg fix
  
  netanel-haber
  authored and
  dcampora
  committed
  7bec668
- Adapt keywords.
  dcampora
  committed
  98a8504
- Update expected output in multimodal test.
  dcampora
  committed
  08db006
- Put multimodal tests on torch sampler due to memory consumption.
  dcampora
  committed
  f66a946
- Waive flaky test_ptp tests.
  dcampora
  committed
  8f678cb

Navigation Menu

Repository navigation

[TRTLLM-6785][feat] BREAKING CHANGE Enable TRTLLM sampler by default - #6216

[TRTLLM-6785][feat] BREAKING CHANGE Enable TRTLLM sampler by default#6216
dcampora merged 32 commits into
NVIDIA:mainfrom
dcampora:user/dcampora/enable_trtllm_sampler_default

Commits on Aug 7, 2025

Make TRTLLM Sampler the default.

Fix default.

Fix test_return_logits.

Wider support of executor_config.speculative_config.

Set max batch size to 64 in test_llm_api_pytorch qwen

Fix run of tests against spec decoding)

Fix max in max_attn_window values.

Formatting.

Remove enable_trtllm_sampler option.

Unwaive qwen test, adapt sampler output.

Adapt e2e test format.

Use Qwen3/Qwen3-0.6B-Base instead of TinyLlama for its longer max_position_embeddings.

Remove unnecessary enable_torch_sampler in test.

Removed unnecessary setting of Torch sampler to false.

Change enable_torch_sampler to use_torch_sampler.

Fix remaining single gpu issues.

Fix remaining failing tests.

Apply test correction.

test_disaggregated_torch_sampler

Use torch sampler for tests that use large models.

refactor: Rename maxBatchSize to maxNumSequences in decoder and gptDecoder classes

finalize_events: dict[str, CudaEvent] | None = None

wording

Adapt to use_torch_sampler.

Fix test_openai_chat_multimodal.

setup is_attention_dp_dummy=True gen requests

Use torch sampler in multimodal.

minimal disagg fix

Adapt keywords.

Update expected output in multimodal test.

Put multimodal tests on torch sampler due to memory consumption.

Waive flaky test_ptp tests.