* sd: sync with master-801-9cfe2af
* sd: sync with master-802-e92e86f
* sd: sync with master-805-e31a86c
* sd: sync with master-810-db99efd
* sd: sync with master-812-ea7f0c8
* sd: minimax-h3 support
In autoswap mode, a POST to /v1/completions or /v1/chat/completions
carrying a `model` name that is not an entry in the admin dir set
`model_switch_pass = True` before checking the whitelist. No swap was
performed, but the flag suppressed the request-type dispatch below it,
so the text model was never loaded on demand.
The same requests without a `model` field, and every other model type
(stt/tts/embed/music/image), skip that branch entirely and load fine --
which is why only chat was affected, and why sending one model-less
request worked around it. It also recurs after --adminunloadtimeout
fires, since the "nomodel" state is recovered from by that same
dispatch.
Only set the flag on the path that actually issues the reload.
The reasoning budget derived from reasoning_effort never applied to Mistral
models. gpttype_adapter.cpp picks the think delimiters from a switch on the
model architecture, and mistral3 has no case, so it falls back to <think> /
</think>. Those are not vocabulary tokens for Ministral-3, so TokenizeString
returns more than one token each, the expected_start/end_tokens guard clears
all three vectors, and apply_reasoning_budget() returns at its first if.
The parameter is accepted, converted and passed down to the sampler, then
dropped on a size check, with nothing logged.
Adding the mistral3 case arms the budget. [THINK] and [/THINK] are single
vocabulary tokens (ids 34 and 35 on Ministral-3), so the size guard passes.
The thinkformats entry is a separate fix for a separate defect: without it the
thinking block was never split out, so it leaked into content with its [THINK]
marker still in it, instead of going to reasoning_content.
Measured on Ministral-3-14B-Reasoning-2512 (IQ4_XS, ctx 8192, --jinja), 5 real
prompts x 3 samples per cell, max_tokens 3000 (so a 750-token budget at "low"):
reasoning_effort thinking words before thinking words after
none 311 - 2314 7 (the forced-close phrase)
low 340 - 2255 521 - 574
Forced closes: 0/15 before, 14/15 after at "low" and 15/15 at "none". Three
samples per cell because this model's variance at temperature 0.7 spans a
factor of 4 on an identical payload — a single sample per cell cannot tell an
effect from noise.
No regression on a non-reasoning mistral3 model: Ministral-3-8B-Instruct with
reasoning_effort "low" returns finish_reason "stop", a normal answer and zero
forced closes, since apply_reasoning_budget() bails out when the start marker
never appears.
When a tool/MCP result carries an image, the OpenAI-compatible chat
adapter's jinja code path left the base64 payload in the rendered
prompt as plain text (a single 1024x1024 jpeg bloated the context by
~120k tokens), while the legacy path already stripped it via
strip_mcpcontent_of_media.
- format_jinja now strips the base64 from tool-role string content
before rendering, matching the legacy path; the image itself is
still swept out and attached separately.
- sweep_media_from_messages now also recognizes MCP-style image
content blocks (type == "image") inside a content list, so images
delivered that way are attached instead of dropped.
* SYCL: add oneMKL GEMM flash attention for XMX-accelerated prompt processing
* fattn-mkl: fix interleaved dst layout in normalize kernel
- Fix mkl_fa_normalize_head: use interleaved dst layout
((query * n_q_heads + head) * DV) matching TILE's
flash_attn_combine_results. Previously used dense head-major
layout which wrote head outputs to wrong addresses, corrupting
attention for all models except Qwen3.6-27B (where GQA=6 heads
were sparse enough to avoid visible overlap).
- Remove 7 redundant stream->wait() calls — SYCL in-order queue
already serializes pure SYCL kernel dependencies. Retain only
the 4 MKL GEMM ↔ SYCL handshake barriers (oneMKL GEMM uses its
own internal queue that does not respect SYCL in-order).
- Remove unused dst_row_stride, diagnostic clutter, and dead
K/V hex dump (fa_diag block in fattn-mkl.cpp).
- Add MKL_FA_DISABLE=1 env var for A/B testing.
- Add FA-DISP watchdog (MKL_FA_DEBUG=1) and FA-DIAG output
fingerprint (MKL_FA_DIAG=1) in fattn.cpp.
Tested: Gemma-4-26B, Gemma-4-31B, Qwen3.6-27B, Qwen3.6-35B-A3B
Perf (B70/Battlemage, 32K, q8_0 KV):
Gemma-4-26B: 1473 t/s MKL vs 746 TILE (1.97x)
Qwen3.6-27B: 609 t/s MKL vs 330 TILE (1.85x)
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* Thank you for the review feedback: rename env vars, use GGML_LOG_INFO, document in SYCL.md
Completed the following:
- Rename MKL_FA_DISABLE → GGML_SYCL_ENABLE_MKL_FA (inverted: 0 to disable)
- Rename MKL_FA_DEBUG → GGML_SYCL_MKL_FA_DEBUG
- Rename MKL_FA_DIAG → GGML_SYCL_MKL_FA_DIAG
- Replace fprintf(stderr, ...) / fflush(stderr) with GGML_LOG_INFO() macro
- Document all three env vars in docs/backend/SYCL.md under Runtime
- Add comment explaining MKL FA activation trigger (flash-attn + quantized
KV cache + batch-size >= 1024 + n_kv >= 1024)
Resolves review feedback from arthw.
Again, thank you!!!
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* Thank you for the review feedback round 2: use ggml_sycl_get_env, remove dup waits, gate perf macros
- Replace raw getenv() with ggml_sycl_get_env() in all 4 env-var checks
(fattn.cpp: GGML_SYCL_ENABLE_MKL_FA, GGML_SYCL_MKL_FA_DEBUG,
GGML_SYCL_MKL_FA_DIAG; fattn-mkl.cpp: GGML_SYCL_MKL_FA_DEBUG)
- Remove duplicated stream->wait() before ev.wait_and_throw() in GEMM
KQ and GEMM VKQ — ev.wait_and_throw() already waits for completion
- Gate MKL_ACCUM macro behind do_print so timing accumulators are
no-ops in normal operation
- Remove redundant MIT/Intel copyright header from fattn-mkl.cpp
- Remove unused #include <cfloat>
- Expand SYCL.md MKL FA docs with step-by-step activation trigger
and example llama-cli command
Again, thank you!!!
Co-Authored-By: Claude Code on DeepSeek-v4-Pro
* fattn-mkl: enable MKL FA for all KV cache types
Remove the quantized-only restriction on MKL activation — the MKL
kernel converts any non-F16 K/V to F16 via to_fp16_sycl before GEMM,
so F16 (default), BF16, and F32 caches all benefit from XMX hardware
acceleration. The type restriction was an unnecessary gate.
Before (F16/BF16 default cache + FA on at 32K prefill): ~356 t/s (TILE path)
After: ~670 t/s (MKL path, matching quantized-cache baseline)
Minimal change: two conditions removed, one comment updated in fattn.cpp.
No kernel or conversion code changes — the dequant pipeline already
covers all types.
* fattn-mkl: rename mkl_disable -> mkl_enable for clarity
* fattn-mkl: refine MKL FA dispatch gates
Three changes:
1. Remove quantized-only restriction - MKL FA activates for all
KV cache types (F16 default, BF16, F32, quantized). The MKL
kernel converts non-F16 K/V via to_fp16_sycl before GEMM.
2. Rename mkl_disable -> mkl_enable to match env var
(GGML_SYCL_ENABLE_MKL_FA).
3. Replace batch-size threshold with Q->ne[1] >= 32 gate.
Keeps TG (Q=1) and MTP drafts (Q=3-8) on VEC path where
fused kernel beats MKL launch overhead. Routes all
multi-token prefill through XMX-accelerated GEMM.
Production data confirms Q patterns: 1-8 TG, 32-127 cache reuse,
128+ full reprocess. At 32K F16/BF16 FA-on: 356 -> 670 t/s.
* ggml-sycl: fix F16 cache + MKL FA multi-turn corruption; add gate guards
Two changes:
1. Always copy F16 K/V to dense row-major buffers before MKL GEMM.
Previously F16 was read in-place with raw tensor strides. During
multi-turn conversations, the accumulated KV cache had different
stride properties than a fresh prefill, producing corrupted outputs.
Now dense F16 gets a fast memcpy; interleaved (Gemma) gets a strided
copy kernel. This matches what the quantized paths already did through
to_fp16_sycl.
2. Gate MKL FA on unsupported op params (max_bias, logit_softcap, batch
dim mismatch) and pathological F16 strides (nb[1] not a multiple of
ne[0]*2). These conditions would previously crash inside the MKL
kernel. Pathological strides (test-only) and ALiBi/softcap fall
through to TILE/VEC which handle them correctly.
The stride check uses modulo rather than equality, so both dense
(nb1 == ne0*2) and interleaved (nb1 == H * ne0*2) pass — all real
models use these layouts. Only test cases with overlapping rows
(nb1=32 or nb1=75 for ne0=40) are blocked.
Thanks to hmscider for the oneDNN FA PR (#25222) which surfaced the
same insight: always normalize inputs to contiguous F16 before GEMM.
Co-Authored-By: Claude Code using DeepSeek-V4-Pro <noreply@anthropic.com>
* fattn-mkl: fix quant+GQA KV strides, tighten MKL gate, add K>=1024 tests
Adding K>=1024 flash-attn test cases surfaced several MKL bugs:
- Quant K/V with a padded seq-view (real KV cache) used the wrong
strides in the dequant path... only the true Gemma interleave
layout should reconstruct strides. nb[2] vs ne[1]*nb[1]
- Gate was firing on shapes the kernel doesn't handle: head_dim < 64
or not a multiple of 64, MHA, attention sinks, and
bf16 decode... fell through to vec which no bf16 case.
Gate MKL to the validated envelope: gqa>=2, head_dim 64 through 512
(has to be a multiple of 64) with matching K/V head size, mask,
no sinks/alibi/softcap... everything else falls back to tile.
Covers Qwen Dense/MoE and Gemma4 Dense/MoE
Ran test-backend-ops -o FLASH_ATTN_EXT: 3641/3641 pass.
Perplexity unchanged... 6.7267 MKL vs 6.7290 stock using
Qwen 27b q5_k_xl
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* fattn-mkl: bound attention scratch so it doesn't grow with batch or context... also dropped the bf16 comment in fattn.cpp per arthw review.
* Update ggml/src/ggml-sycl/fattn-mkl.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* Update ggml/src/ggml-sycl/fattn-mkl.cpp
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* apply arthw suggestions: enum for dequant modes, macro for wg_size, env-var one-liners
---------
Co-authored-by: Claude Code using DeepSeek-V4-Pro <noreply@anthropic.com>
Co-authored-by: Neo Zhang <zhang.jianyu@outlook.com>
* llama : enforce the same K and V cache types for DeepSeek V4; enable FA if V cache is quantized
* llama : enforce the same K and V cache types for MLA models
---------
Co-authored-by: Stanisław Szymczyk <sszymczy@gmail.com>
* improve fa of quantized kv cache
* Fix some bugs and some comments.
* fix v type check and some comments
* Fix build error caused by rebasing
* editorconfig checking pass
* add bool cwhn = true to conv_2d test cases
* add layout check at graph building time
* extend layout checks for conv2d.cu kernel
* in CPU back-end kernel needs to be stored contiguously to prevent test failures with cwhn=1
* trim white space
* do op support check in vulkan backend
* fix CI failure and vulkan run-time assert failure by introducing new graph build-time check in ggml_backend_vk_device_supports_op
* add additional check in support_op function for Vulkan to fix run-time assert failure
* metal: fix memory leak if model is freed without any GPU operations
* metal: run dummy work only if residency sets are used
* metal: wrap function in #if defined
* metal: measure system-wide wired memory in test
* metal: always build regression test
Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>
---------
Co-authored-by: YiChen Lv <63285796+forforever73@users.noreply.github.com>
* fix: single-flight conversations store init
* refactor: remove unused legacy-migration util
* fix: make createSystemMessage transactional
* fix: delete message branches cascading on edit/regenerate
* fix: stop stamping lastModified on conversation metadata updates
* fix: count cascaded forks in bulk delete toast, bulkify deleteAll
* refactor: drop redundant conversation list respreads
* refactor: create conversation in a single write
* fix: use table constant in toggleConversationPin
* fix: keep the system message placeholder out of the edit form
* fix: keep focus in the system message editor after opening it
* fix: focus the main chat form after submitting a system message
* fix: update timestamp of the correct conversation on stream completion
ggml_cuda_should_use_mmq() selects MMQ purely from the quantization
type. The current MMQ configurations are designed and maintained against
a minimum of 48 KiB per-block shared memory, the limit provided by
NVIDIA Pascal GPUs and later. On devices that report less, no supported
MMQ tile fits and mul_mat_q_switch_J() aborts when every tile size
exceeds the device's per-block shared memory budget.
Disable MMQ when smpbo < 48 KiB so the caller falls back to the BLAS
path instead of hitting GGML_ABORT. Some current MUSA QY1 devices
report only 28 KiB and are covered by this guard.
Reproduced on a Moore Threads MTT S70 (arch mp_21, 28 KiB shared memory
per block) with an RWKV-7 0.1B Q8_0 model:
$ llama-bench -m rwkv7-g1d-0.1b-Q8_0.gguf -p 128 -n 0
J_best=0
ggml/src/ggml-cuda/template-instances/../mmq.cuh:1521: fatal error
(core dumped)
Only prefill (batch > 1) is affected; token generation is fine. After
the fix the same device falls back to the BLAS path:
Q8_0 pp128 1470.7 t/s, tg8 55.3 t/s (was: abort)
FP16 unchanged
Q4_K_M unchanged
This matches a -DGGML_CUDA_FORCE_CUBLAS=ON build (pp128 1464.2 t/s),
which confirms the fallback path is the one being taken.
This is not MUSA-specific: any device with less than 48 KiB per-block
shared memory is affected.
Co-authored-by: KakaruHayate <KakaruHayate@users.noreply.github.com>
Adds trace logging in server-context.cpp for slot similarity checking
during prompt cache slot selection, including skip reasons and similarity
calculation details.
Assisted-by: llama.cpp:Qwen3.6-27B
Concedo
Wagner Bruna
Tai An
Julien BODIN
Recoordinate
Ruixiang Wang
David Friehs
Ozymandias_EBON
Neo Zhang
fairydreaming
Sachin Sharma
Titaniumtown
Masashi Yoshimura
Xuan-Son Nguyen
Jeff Bolz
pmaybank
o7si
Georgi Gerganov
Robert Esclapez
Pasha Khosravi
timkhronos
Niklas Wenzel
Aleksander Grygier
Jonathan Clohessy
Tunahan
Kakaru
Alessandro de Oliveira Faria (A.K.A.CABELO)
Kaben Nanlohy