memory: respect unified KV cache in hybrid memory for eval tasks#21224
Merged
ggerganov merged 1 commit intoggml-org:masterfrom Apr 1, 2026
Merged
memory: respect unified KV cache in hybrid memory for eval tasks#21224ggerganov merged 1 commit intoggml-org:masterfrom
ggerganov merged 1 commit intoggml-org:masterfrom
Conversation
The hybrid memory paths (`llama-memory-hybrid.cpp` and `llama-memory-hybrid-iswa.cpp`) always used sequential equal split, ignoring the unified KV cache flag. This caused hellaswag, winogrande, and multiple-choice evaluations to fail on hybrid models (models with both attention and recurrent/SSM layers, such as Qwen3.5-35B-A3B) with: split_equal: sequential split is not supported when there are coupled sequences in the input batch (you may need to use the -kvu flag) PR ggml-org#19954 fixed this for `llama-kv-cache-iswa.cpp` by automatically enabling unified KV mode and setting n_parallel >= 4 for multi-choice eval tasks. However, the hybrid memory paths were not updated. This commit mirrors the iswa fix: use non-sequential split when KV cache is unified (n_stream == 1), which is automatically set by llama-perplexity for hellaswag/winogrande/multiple-choice since ggml-org#19954. Tested on Qwen3.5-35B-A3B (hybrid attention+SSM MoE model): - HellaSwag: 83.0% (400 tasks) - Winogrande: 74.5% (400 tasks) - MMLU: 41.2% - ARC-Challenge: 56.2% - TruthfulQA: 37.7% All previously failed with llama_decode() error.
ggerganov
approved these changes
Mar 31, 2026
ggerganov
added
the
merge ready
icex
added a commit
to icex/llama.cpp
that referenced
this pull request
Apr 5, 2026
Includes: - fix: handle non-capturing groups (?:...) in JSON schema pattern converter (ggml-org#21124) - memory: respect unified KV cache in hybrid memory for eval tasks (ggml-org#21224) - fix: CUDA FA kernel selection, head dimension 512 support - rotate activations for better quantization (ggml-org#21038) - Various parser, jinja, webui, and CI fixes Conflicts resolved: - llama-kv-cache.cpp: keep TurboQuant InnerQ stubs + upstream Hadamard helpers - llama-graph.cpp: keep TurboQuant V-padding + upstream self_v_rot - fattn-tile.cu: add upstream D=512 before TurboQuant HIP guard - fattn.cu: combine D=512 (upstream) + D=640 (TurboQuant) exclusions
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
The hybrid memory paths (
llama-memory-hybrid.cppandllama-memory-hybrid-iswa.cpp) always used sequential equal split, ignoring the unified KV cache flag. This caused hellaswag, winogrande, and multiple-choice evaluations to fail on hybrid models (models with both attention and recurrent/SSM layers, such as Qwen3.5-35B-A3B) with:PR #19954 fixed this for
llama-kv-cache-iswa.cppby automatically enabling unified KV mode and setting n_parallel >= 4 for multi-choice eval tasks. However, the hybrid memory paths were not updated.This commit mirrors the iswa fix: use non-sequential split when KV cache is unified (n_stream == 1), which is automatically set by llama-perplexity for hellaswag/winogrande/multiple-choice since #19954.
Tested on Qwen3.5-35B-A3B (hybrid attention+SSM MoE model):
Additional information
Requirements