Skip to content

Main doc fix - #20

Merged
juney-nvidia merged 1 commit into
mainfrom
main_doc_fix
Oct 20, 2023
Merged

juney-nvidia merged 1 commit into
mainfrom
main_doc_fix

Conversation

@juney-nvidia

Copy link
Copy Markdown
Collaborator

No description provided.

@juney-nvidia
juney-nvidia changed the base branch from release/0.5.0 to main October 20, 2023 00:49
@juney-nvidia
juney-nvidia merged commit 103bf69 into main Oct 20, 2023
@juney-nvidia
juney-nvidia deleted the main_doc_fix branch October 20, 2023 00:52
wu1du2 pushed a commit to wu1du2/TensorRT-LLM that referenced this pull request May 11, 2025
lancelly added a commit to lancelly/TensorRT-LLM that referenced this pull request Aug 9, 2026
…d tokens would have been accepted

Direct counterfactual measurement (cap-accept mode: same planner
decisions, full-block verify, acceptance capped): poetry trims 2.08
positions per request per step of which 6.2% were alive (0.130 tokens,
9.8% of no-trim acceptance); arena trims 0.71 of which 16.1% were alive
(4.3%). The decision layer is exonerated -- the confidence screen works
and the planner correctly trims less on the high-accept workload. This
closes the responsibility chain: decisions good, table economics sound
at true 128-row steps, execution (batch fragmentation into 64-row
floor-dominated steps) is the sole failure point. Also corrects the
SGLang profiler characterization: it supports a 2D additive table via
--fracs (the lineage of our NVIDIA#26 frac sweep); the canonical/blog flow and
both cross-stack runs used the 1D static mode.

Signed-off-by: Lance Liao <108499334+lancelly@users.noreply.github.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant