Repository navigation
add SIGINT support for _WIN32 environments - #120
Merged
Merged
Conversation
ggerganov
approved these changes
Mar 15, 2023
blackhole89
pushed a commit
that referenced
this pull request
Mar 15, 2023
* add SIGINT support for _WIN32 environments * perhaps more consistent
jesusmb1995
pushed a commit
to jesusmb1995/llama.cpp
that referenced
this pull request
Apr 15, 2026
ggml_opencl: add on-disk kernel binary cache for faster startup
Seunghhon
pushed a commit
to Seunghhon/llama.cpp
that referenced
this pull request
Apr 26, 2026
* add SIGINT support for _WIN32 environments * perhaps more consistent
phuongncn
pushed a commit
to phuongncn/llama.cpp-gx10-dgx-sparks-deepseekv4
that referenced
this pull request
Apr 28, 2026
* Adding q8_0_r4 We get PP-512(LLaMA-3.1-8B) = 268 t/s on a Ryzen-7950X compared to 175.6 t/s for Q8_0. * q8_0_r4: NEON We get PP-512(LLaMA-3.1-8B) = 112.6 t/s on M2-Max. * q8_0_r4: Zen4 matrix-vector specialization --------- Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
ljubomirj
pushed a commit
to ljubomirj/llama.cpp
that referenced
this pull request
May 6, 2026
* add SIGINT support for _WIN32 environments * perhaps more consistent
Soket1
pushed a commit
to Soket1/llama.cpp-xdna
that referenced
this pull request
May 29, 2026
The xdna_layer_fused_sizes allocator mirrored the (buggy) IRON
formula for total_q / total_kv1 that omits the `/ m_input_qkv`
divisor. The host packer used the correct per-tile formula, so:
- bo0 was allocated 2x oversized for Q + each KV region
- packer wrote Q at byte 4096, K at byte 2363392, V at byte 2952216
- design.py read Q at byte 4096 (matches), K at byte 4722688
(past packer, in zero-tail), V at byte 5312512 (likewise)
Net effect: Q phase produced finite output, K and V phases produced
all-zero output (the long-standing "K/V phase block" attributed in
prior memos to a colN_body wiring issue is actually this).
Sync the C++ allocator with the IRON-side fix (commit
IRON-windows@ba67dc6). After fix:
bo0_bytes 7081984 -> 3543040 (Llama-3.2-1B)
q_rot fin=2048/2048, k_rot fin=512/512, v fin=512/512 (was K=V=0)
Also:
- Bump layer_fused cache key v1 -> v2 to invalidate stale xclbins
(cache key bumped in both compile.py and ggml-xdna.cpp per
`Bump BOTH Python and C++ cache keys` rule).
- Add P0.3 BUG-FIX VALIDATION probe: one-shot read of bo4 at
off_q_rot / off_k_rot / off_v, count finite bf16 elements.
outL probe stays for the Stage 3 drain-wiring work.
outL is still all-zero — design.py drains only Q/K/V; O_proj +
attention + FFN have no drains yet (P0.4, task ggml-org#120).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Soket1
pushed a commit
to Soket1/llama.cpp-xdna
that referenced
this pull request
Jul 26, 2026
Both quickstarts predated this fork and pointed elsewhere: clone commands for albiol2004/llama.cpp, IRON from amd/IRON (which lacks the ~70 operators this backend dispatches, so the first NPU op ImportErrors), and XRT downloads from amd/xdna-driver/releases, where there are no releases at all. The numbers came from the same era -- "decode ~9 t/s", a model table listing Gemma 3 1B and Llama 3 8B as supported. Rewritten around the measured envelope instead. The fast path is one fused decode kernel built for one geometry, and the docs now say so up front: XDNA 2 only (compile.py hardcodes 8 columns on Windows, so XDNA 1 is not handled), Llama-3.2-1B, Q4_0. Anything else degrades to per-op dispatch or CPU. Numbers re-measured today on a Ryzen AI 9 365 rather than carried over: decode 10.6 -> 29.0 t/s, but prefill 191.6 -> 100.3. The prefill regression is stated plainly -- this wins at generation and loses at prompt processing, and a doc that hides that sends people to the wrong workload. Two traps that cost real debugging time are now written down: XILINX_XRT is configure-time only (left set at runtime the kernel compile dies with no output), and -fa off is required (with -fa on the fused layer never dispatches and you silently get the slow path). Both verified by running them. The kernel cache now ships in the repo, so the run instructions no longer imply a toolchain that is not needed for the supported model. XDNA_TRACKER.md and XDNA_OPTIMIZATION_PLAN.md are marked as historical rather than edited: both describe the ~9 t/s baseline and read as current roadmaps. Co-Authored-By: Claude Opus 5 <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
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.
No description provided.