ngram-mod: Reset i_last when low acceptance streak occurs#22168
Merged
Conversation
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
Member
|
What speculative decoding parameters did you use in this test? |
Contributor
Author
|
Oh, sorry, it appears in cleaning up the layout I deleted the lines with the parameters. Additionally, I had to disable reasoning ( I've been using this change ever since proposing it for agentic coding with slightly different parameters: In my build I also added a little debug logging, and it is not uncommon to see a low acceptance streak right before a big chunk of 100% acceptance. With i_last not being reset, it wouldn't have anything to work with in those cases. |
ggerganov
approved these changes
Apr 21, 2026
TheTom
added a commit
to TheTom/llama-cpp-turboquant
that referenced
this pull request
Apr 22, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
TheTom
added a commit
to TheTom/llama-cpp-turboquant
that referenced
this pull request
Apr 22, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
arthw
pushed a commit
to arthw/llama.cpp
that referenced
this pull request
Apr 23, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
4 tasks
rsenthilkumar6
pushed a commit
to rsenthilkumar6/llama.cpp
that referenced
this pull request
May 1, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
samuraieng
pushed a commit
to samuraieng/llama.cpp
that referenced
this pull request
May 6, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
ljubomirj
pushed a commit
to ljubomirj/llama.cpp
that referenced
this pull request
May 6, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
meh
pushed a commit
to meh/llama.cpp
that referenced
this pull request
May 10, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
meh
pushed a commit
to meh/llama.cpp
that referenced
this pull request
May 10, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Jcfunk
pushed a commit
to Jcfunk/llama.cpp
that referenced
this pull request
May 13, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Jcfunk
pushed a commit
to Jcfunk/llama.cpp
that referenced
this pull request
May 13, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
my-other-github-account
pushed a commit
to my-other-github-account/llama.cpp
that referenced
this pull request
May 15, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
my-other-github-account
pushed a commit
to my-other-github-account/llama.cpp
that referenced
this pull request
May 15, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
8 tasks
baramofme
pushed a commit
to baramofme/llama-cpp-turboquant
that referenced
this pull request
May 23, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
fewtarius
pushed a commit
to fewtarius/llama.cpp
that referenced
this pull request
May 30, 2026
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
wel97459
pushed a commit
to wel97459/llama-cpp-turboquant
that referenced
this pull request
Jun 4, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
wel97459
pushed a commit
to wel97459/llama-cpp-turboquant
that referenced
this pull request
Jun 4, 2026
Cherry-picks 4 upstream PRs to enable speculative decoding on hybrid MoE+SSM architectures (Qwen3.6-35B-A3B): - ggml-org#19493 — speculative checkpointing (save/restore recurrent state) - ggml-org#22114 — refactor "use checkpoint" logic - ggml-org#22168 — reset i_last on low acceptance streak - ggml-org#22223 — add --spec-default argument Smoke tested on M5 Max with turbo4 KV — zero regression. Co-Authored-By: tturney@psyguard.ai Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Overview
By resetting i_last to zero, we will include the current context when rebuilding the speculative map.
The existing behavior would skip the current context, thereby often losing the benefit of speculative decoding during later parts of the generation.
The effect of this seems to depend on both the model and the speculation parameters.
Benchmark:
vllm bench serve --model google/gemma-4-26B-A4B-it --host 127.0.0.1 --port 9876 --num-prompts 96 --dataset-name hf --dataset-path vdaita/edit_5k_char --backend openai-chat --endpoint '/v1/chat/completions' --max-concurrency 1Gemma 4 26B-A4B:
Qwen 3.6 35B-A3B:
As we can see, the effect isn't huge but at 3% to 9% it is still measurable.
The price we pay for it is 1 line of code and a moment longer for speculative map repopulation whenever a low acceptance streak occours.
Requirements