When llama_set_causal_attn(false) is called on a causal model (e.g.
Gemma-4 during vision image decode), llama_set_inputs took the non-causal
else-branch (designed for pure embedding models).
That path wrote the F16 mask with stride n_tokens instead of n_kv, and iterated batch
indices rather than KV cache cells.
The result was that every image query row beyond the first was
written at the wrong offset, leaving stale -inf values from
previous decodes visible to the GPU kernel. Any conversation
that had built up prior KV mask data would produce all-inf attention scores
for most image tokens, collapsing softmax to NaN and aborting at sampling.
Resolves#1984
* (qwen3vl) Correct calculation for injection point of deepstack image embeddings
INjection point for deepstack embeddings used Hyperparameter n_embd_inp(), which caused the hidden state to be double accounted for, causing an OOB array access. The correct accessor is n_embd()
* Fix m-rope when pipeline parallelism is enabled
* wip: port MTP architecture
Ports the Multi-Token Prediction (MTP) architecture to the older `llama.cpp` codebase used by `ikllama`.
Changes include:
- Updating `llama_batch` to support `mtp_params`.
- Modifying `llama_decode_internal` (and `encode`) to handle MTP operations (Warmup, Update, Draft).
- Adding public APIs for MTP state management (`llama_set_draft_input_hidden_state`).
- Adapting the embedding extraction logic to skip MTP update passes.
* Refactors `server_slot` to support generic speculative decoding (MTP or Draft Model).
* core: enable hybrid outputs (logits + embeddings) for MTP support
* fix(mtp): correct KV-cache slot finding for updates
* fix(mtp): persist hidden states to prevent context corruption during drafting
* refactor(mtp): clean unused code
* fix(mtp): update server to new functions name
* fix(mtp): fix graph and save hidden state
* mtp: refactor integration, context params and kv cache search
* mtp: fix hidden state extraction and speculative acceptance flow
* server: fix MTP warmup for long prompts and reset token buffer
* llama: refactor MTP operation state to context parameters
* server: fix n_past calculation in MTP acceptance
* llama: fix mtp enable flags
* speculative: refactor MTP to use common_speculative interface
* context: remove unused signatures
* clip: fix deprecated enum-enum conversion warning
* common: fix format string crash in help message
* context: fix mtp activation logic
* llamat: always use the extracted embedding
* llama: get all embeddings to kv cache
* llama: revert logit to not run mtp for not supported arch
* llama: allocate all the n_outputs for MTP
* wip
* server-context: get only the last embedding for hidden state
* ggml-backend: fix array of bounds in debug build
* server-context: run mt kv update to each prompt batch
* revert segmentation fault fixes
* glm-mtp(feat): optimize graph embedding and recursive drafting
* glm5-mtp(feat): add glm 5 mtp logic
* glm-mtp: standardize the MTP graph
* glm 5 mtp: apply post-layer cvec
* glm 5 mtp: mark head as mandatory
* get normed embeddings for glm 5
* Fix GLM5 MTP
* GLM5 MTP: just reuse the layer attention implementation
* Make MTP work with split mode graph
---------
Co-authored-by: samuel <samueloliveira32df@gmail.com>
* Disable K Hadamard transform if K-head size is not a power of 2
* Allow Hadamard transform for head sizes that are not power of 2
* Give more details why Hadamard is not possible
* Arghh
llm_load_tensors stores `default_layer_device[i]` as a local index into
`model.devices` (consistent with `device_mem[]`, `model.splits[]`, and
all graph-building consumers), but the four
`llama_default_buffer_type_offload(model, default_layer_device[i])`
callsites passed it through as if it were a raw post-CVD device id.
Under `-dev`/`-devd` subsets where `model.devices != {0..N-1}`, this
selected the wrong buffer type. Wrap with `model.devices[...]` to match
the existing `model.devices[main_gpu]` pattern on the adjacent lines.
llama_init_from_model has the same bug for `main_gpu`: every consumer
(auto-fit override at line 3428, MTP clamp, the `model.devices[main_gpu]`
translations at lines 3678/3682, and graph-building `splits[main_gpu]`)
treats it as a local index, but the five single-GPU backend init paths
(CUDA, Vulkan, SYCL, Kompute, CANN) pass `model->main_gpu` straight to
the backend init, which expects a raw device id. e.g. `-dev CUDA1` with
default `--main-gpu 0` and `split_mode=NONE` called
`ggml_backend_cuda_init(0)` instead of `cuda_init(1)`. Compute
`main_gpu_id` once and use it for all five paths.
* ggml: ggml_dequant_hadamard fused op for MLA -khad path
Adds a new ggml op that fuses (ggml_cast -> F32) + (ggml_hadamard) into a
single kernel. Reads a quantized (or F16/F32) source and produces a per-
Hadamard-block F32 chunk with the inverse transform applied, without
materializing a full-size F32 intermediate buffer.
Motivation: the MLA pp_opt path in build_deepseek2.cpp un-encodes the
H-applied cache_nope view at every PP call. Today that runs as a cast
(quant -> F32) followed by a separate ggml_hadamard kernel, costing two
full-size F32 passes per layer per rank per call. Fusing them halves
the bandwidth on the un-encode and removes one kernel launch.
CUDA kernels in dequant_hadamard.cu lift the Walsh-Hadamard butterfly
from hadamard.cu and dequant helpers from dequantize.cuh:
* qr=1 layout (q8_0): consecutive dequant pair, stage 1 fused with load
* qr=2 layout (q4_0 / q4_1 / q5_0 / q5_1 / q6_0 / iq4_nl): dequant pair
at stride qk/2, explicit stage 1 after sync
* F16 has a dedicated kernel
* F32 source falls back to the standalone Hadamard op
CPU impl in iqk_cpu_ops.cpp composes the existing type_traits.to_float
dequant with fast_ht for graph completeness. nh in {64, 128, 256, 512}.
* MLA-TP: Hadamard pretransform of wv_b/wk_b_pp for -khad
Fold the 64-block orthonormal Hadamard into wv_b and wk_b_pp once at
context init so the pp_opt mul_mats consume the K cache in its on-disk
encoded basis. The per-PP-call cache_nope un-Hadamard is then skipped
(rope half still un-applied — it goes to FA via concat, no wk_b multiply).
Math is identity by H^T H = I: mul_mat(H@wv_b, H@cache) = wv_b^T @ cache.
For mla=2/3 absorb, composes correctly with the existing post-FA
ggml_hadamard(kqv_compressed, 64).
All-or-nothing across layers under a castable type-allowlist (excludes
1-3 bpw IQ types whose requant blows up beyond PPL noise). Models with
ineligible weights fall back to the runtime un-Hadamard path unchanged.
Composes with the fused ggml_dequant_hadamard op (prior commit): with the
fold active only the rope half still runs the runtime transform, via the
fused kernel.
* MLA-TP: fix TG with -khad after wv_b/wk_b_pp fold
The absorb branch of build_deepseek2_tp_attention applies
ggml_hadamard to kqv_compressed after FA, then multiplies by
wv_b. Pre-fold this was needed because wv_b was un-encoded; with
the wv_b fold (prior commit) the mul_mat already expects
H-encoded kqv_compressed:
mul_mat(H @ wv_b, kqv_encoded) = wv_b^T @ H @ H @ kqv_unencoded
= wv_b^T @ kqv_unencoded (H @ H = I)
Skip the post-FA hadamard when model.khad_pretransformed is set
so the two H applications cancel instead of double-applying.
Affects the absorb branch: TG (n_tokens=1), short-context PP
(n_kv < 1024), and models without wk_b_pp. Long-context PP goes
through the pp_opt branch and is unrelated/unchanged.
Reported by @ikawrakow on PR 1852. Verified across mla={1,2,3} x
khad={on,off} x -ctk={q8_0,q4_0} on GLM-4.7-Flash IQ5_K and the
unsloth IQ4_XS variant ik used to reproduce.
* ggml_hadamard: accept F16 and quant sources; drop GGML_OP_DEQUANT_HADAMARD
Per @ikawrakow review on PR 1852: subsume the per-source-type dispatch
into the existing GGML_OP_HADAMARD instead of carrying a separate enum
entry, op constructor, and standalone files.
ggml_hadamard's API is unchanged from the call-site perspective. The
constructor's F32-only assertion is dropped; ggml_cuda_op_hadamard and
iqk_hadamard now dispatch internally:
- F32 source: existing F32 butterfly (unchanged)
- F16 source: dedicated kernel
- q8_0 / q4_0 / q4_1 / q5_0 / q5_1 / q6_0 / iq4_nl: fused dequant +
butterfly kernel (lifted from the deleted dequant_hadamard.cu)
- CPU side composes traits.to_float with fast_ht
Net diff: -80 lines. Removes dequant_hadamard.{cu,cuh}, the enum entry,
op table rows, ggml_dequant_hadamard constructor, dispatch cases, and
the DEQUANT_HADAMARD supports_op block.
Verified clean build + TG smoke (mla=3 +khad q8 on GLM-4.7-Flash-IQ4_XS,
same coherent output as prior commit on feat/dequant-hadamard).
* MLA TP prompt processing optimisation
Adds a per-rank prompt-processing path to build_deepseek2_tp_attention
that materialises K/V from the compressed latent cache and runs a
standard flash_attn instead of the FlashMLA-3 absorb kernel the TP
attention currently uses for all batch sizes. Affects MLA archs under
-sm graph (DEEPSEEK2, GLM_DSA, MISTRAL4).
Gated on n_tokens >= 128 (set by caller) AND n_kv >= 1024. Below
either threshold the absorb path runs unchanged. Token generation
takes the absorb path; only prompt processing at non-trivial context
materialises.
A second piece pre-computes wk_b in a pp_opt-favouring orientation
(wk_b_pp: [kv_lora_rank, qk_nope, n_head]) at llm_prepare_mla time,
so the per-PP-call materialise can mul_mat against the latent cache
directly without an F16 cast + permute + ggml_cont on wk_b each call.
Path A (wkv_b in GGUF) and Path B (only wk_b/wv_b in GGUF) both
populate wk_b_pp through the standard per-rank replica setup.
Measured on 8x RTX 3090, -sm graph -mla 2 -fa on:
DSV2.5 IQ2_XS c=8k ub=2048 PP +51% to +60%
GLM-4.7-Flash IQ4_XS c=32k ub=2048 PP -6% (PP@0) to +77% (PP@30720)
GLM-5.1 IQ1_S q4_0 c=16k ub=2048 PP +5% to +9%
PPL parity within +/-0.2 noise (DSV2.5 bit-identical 5.3917, GLM-4.7
8.83 vs 8.96, GLM-5.1 6.96 vs 7.00). Token-generation throughput
unchanged within noise.
Compute buffer at init:
DSV2.5 -54 MiB total (allocator noise)
GLM-4.7-Flash +1042 MiB total (~+173 MiB per non-output device)
GLM-5.1 0 (MoE intermediates dominate)
* MLA TP: respect mla=1 vs mla=3 distinction, rename attn_k_b_pp -> attn_kv_b
ikawrakow/ik_llama.cpp#1841 review feedback: the pp_opt path lost the
intended trade-off where mla=1 forgoes pp_opt to save VRAM and mla=3 pays
the wk_b_pp tensor cost for faster long-context PP.
- llm_prepare_mla second pass: gate wk_b_pp synthesis on mla > 1.
Models that ship wk_b in their GGUF (mainline format) no longer
allocate the pp_opt-favoring K weight under mla=1.
- llm_prepare_mla first pass (wk_b synthesis from wkv_b): keep
unconditional under -sm graph. The wk_b_pp materialization here
shares the wk_b_f32 intermediate with the wk_b synthesis above, and
isolating just the wk_b_pp branch leaves the synthesized wk_b in a
state that makes the absorb path produce inf on some quant combos
(DSV2.5 IQ2_XS). Trade: the synthesized-wkv_b path still pays the
wk_b_pp allocation under mla=1, but the bigger compute-buffer
saving (no pp_opt branch at runtime) still applies.
- build_deepseek2 outer pp_opt: include cparams.mla_attn > 1 in the
pp_opt definition itself, so mla=1 is bypassed throughout (TP and
non-TP attention paths).
- build_deepseek2 tp pp_opt: require wk_b_pp present. Drop the dead
runtime wk_b transpose fallback (unreachable now that wk_b_pp is
guaranteed when tp_pp_opt fires).
- llama_kv_cache_init: have_wkv_b probe now treats wk_b_pp (attn_kv_b)
as equivalent to wkv_b for the purposes of allowing mla>1 to stay
put. Without this, -sm graph models that have wk_b/wv_b separately
in the GGUF (no combined wkv_b) would silently downgrade to mla=1.
- Rename the synthesized tensor "attn_k_b_pp.weight" -> "attn_kv_b.weight"
to match the mainline naming ik uses.
GLM-5.1 in particular benefits: its mla=3 PP improvement over mla=1 is
negligible on this arch (~0.4% in our sweeps), so users save the
runtime cost by sticking to mla=1.
* MLA tensor parallelism under -sm graph (DEEPSEEK2/GLM_DSA/MISTRAL4)
Extends -sm graph (split-mode graph) to MLA-style attention across the
DEEPSEEK2, GLM_DSA, and MISTRAL4 architectures. Previously these archs
fell back to -sm layer regardless of the user's flag.
Implementation:
- Per-rank attention build in build_deepseek2_tp_attention with
view-sliced FlashAttention, split-buffer output projection, and
ggml_reduce across devices
- wk_b / wv_b absorbed weights replicated per device via materialize()
in llm_prepare_mla (these can't live in a split buffer)
- KV cache replication path (replicated_k_l) for graph-mode TP
- distribute_mla_tensors_for_split_mode_graph routes attention/norm
tensors into ctx_split; expert tensors stay per-layer
- Implements ggml_backend_cuda_split_buffer_get_tensor for the
replicated / row-split / col-split inverse paths
- Early-reject guard in src/llama.cpp that auto-downgrades -sm graph
to -sm layer (with a warning) when incompatible loader flags are set:
-ncmoe, -cmoe, -ot, -rtr, -muge
New CLI flag:
- -gap | --graph-attn-precision <f16|f32> (default f16)
See the PR description for the full validation matrix (3 archs x 2/4/8
GPU counts), perf numbers, VRAM accounting, and known limitations.
* Some tweaks
* materialize lambda: per-head split for graph-mode tp_replicate
7dd19e19 changed wk_b/wv_b distribution from mirror to per-head split
(split_dim=2) via prepare_split_tensors. That path only fires when
wk_b/wv_b are loaded from GGUF.
Models that store only wkv_b in GGUF derive wk_b/wv_b at load via
llm_prepare_mla, going through the materialize lambda, which was
untouched and still produced mirror replicas (split_dim=-1, full n_head
per device).
build_deepseek2_tp_attention now does mul_mat(wk_b_local, q_nope_perm)
without the prior view_3d slice, so a mirror replica passes an n_head
tensor where the kernel expects n_head_local. Result: silent SIGSEGV
right after model load.
Mirror logic in materialize is replaced with the same per-head split as
prepare_split_tensors: head_offsets derived from wo split, each rank
gets a tensor with ne[2]=n_head_local, data copied from the appropriate
source byte slice. Singular `computed` tensor keeps full metadata for
tensors_by_name lookups.
Tested: 8x3090, -sm graph -mla 3 -fa on now boots cleanly and
sweep-benches without crash. Log confirms new path: "Computed
blk.X.attn_k_b.weight ... split across N devices on dim=2".
* cleanup: indent fix + remove dead view_3d slicing and debug printf
- build_deepseek2.cpp: re-indent the self_attention block in
build_deepseek2_layer_attention (lines 253-670). Block was at column 0
inside a function body; now at the expected 4/8-space indent.
- build_deepseek2.cpp: drop the commented-out view_3d slicing and debug
printfs left over after 7dd19e19's switch to direct mul_mat on
per-rank wk_b_local / wv_b_local. Update the stale 'wk_b is
replicated (split_dim=-1)' comment to match the new split_dim=2
reality.
- ggml-cuda.cu: remove the leftover debug printf in
ggml_backend_cuda_split_buffer_get_tensor.
No behavior change. Verified with a clean rebuild and DSV2.5 +
GLM-4.7-Flash sweep-bench runs.
* llm_load_tensors: gate incompatible-flag warning to MLA archs
The -ncmoe / -rtr / -muge / -ot warning under -sm graph currently fires
for all archs that support graph mode. That's an over-reach: the
incompatibility is specific to the MLA TP paths (DEEPSEEK2, GLM_DSA,
MISTRAL4) — Gemma4 graph mode existed pre-PR and works with those flags.
Gate the warning to MLA archs only.
Also refreshes two stale comments left over from the wk_b/wv_b
mirror -> per-head-split rewrite:
- src/llama.cpp llm_prepare_mla: "Replicate wk_b/wv_b ..." now reads
"Per-head split wk_b/wv_b ..." to match what the materialize lambda
actually does post-823a39e2.
- src/llama-load-tensors.cpp distribute_mla_tensors_for_split_mode_graph:
drop the wkv_b row-split mention (wkv_b is no longer created under
graph mode after 7dd19e19) and correct the wk_b/wv_b distribution
description (per-head split, not per-device replicated).
---------
Co-authored-by: Kawrakow <iwankawrakow@gmail.com>
