Commit Graph
1486 Commits
Author SHA1 Message Date
Lumpiasty 98c7e109b7 llama: evict recurrent/SSM state on device release
The recurrent (SSM/conv) state of hybrid models (e.g. Qwen3.5) was left
resident when a model's device buffers were released for on-demand VRAM
sharing - llama_memory_recurrent::release_device_buffers() was a no-op
default. Implement it (and restore_device_buffers) with the same
capture-host-shadow / free / reallocate pattern as llama_kv_cache, so
llama_memory_hybrid now evicts both its attention KV and its recurrent
state. The state is read-write, so its shadow is recaptured on every
release.

Assisted-by: Claude
2026-11-21 19:39:51 +01:00
Lumpiasty bc55fb9990 server: also free the compute-graph scheduler on KV eviction
Under LLAMA_SLEEP_EVICT_KV a cold model still held the scheduler's worst-case
compute buffer (hundreds of MiB, e.g. ~700 MiB for gemma-26B) plus the resident
experts. release_device(evict_kv) now also frees the sched (sched.reset() +
sched_need_reserve), and restore_device() rebuilds it via sched_reserve(), so a
fully-evicted model holds essentially no VRAM (gemma cold: 7147 -> 51 MiB).
Trades a heavier re-warm (weights+KV H2D + sched reserve) for the freed VRAM;
weights-only mode is unchanged (fast switch, keeps KV+compute).

Assisted-by: Claude
2026-11-21 19:39:46 +01:00
Lumpiasty 8e575c9e54 llama: guard buffer iteration against released device buffers
release_device_weights()/release_device_buffers() leave a null buffer in
ctxs_bufs while the device memory is released for on-demand VRAM sharing.
The memory-breakdown / total_size / clear paths iterated these and called
ggml_backend_buffer_get_size()/get_type()/clear() on the null buffer, tripping
GGML_ASSERT(buffer) and aborting on shutdown of a cold model. Skip null buffers
in llama_model::memory_breakdown and llama_kv_cache::{memory_breakdown,
total_size,clear}.

Assisted-by: Claude
2026-11-21 19:39:46 +01:00
Lumpiasty cb471f9b2f server: optionally evict the KV cache too (Phase 2 of VRAM sharing)
Extend on-demand device residency to the KV cache so that when a model's KV
plus another model would not fit in VRAM, the KV can also be evicted to a host
shadow (D2H on release, H2D on restore) instead of only the weights.

- llama_memory_i: add release_device_buffers()/restore_device_buffers()
  (default no-op). Implemented in llama_kv_cache (D2H shadow of the live
  ctxs_bufs, freed and reallocated like the weights); llama_memory_hybrid and
  llama_kv_cache_iswa delegate to their child caches.
- llama_context::release_device(evict_kv): also evict the memory's device
  buffers when requested; restore_device() rebuilds them. Public API
  llama_context_release_device gains an evict_kv flag.
- server: LLAMA_SLEEP_EVICT_KV=1 enables it. Off by default (weights-only),
  since the KV shadow adds a D2H/H2D copy of the live cache each cycle.

Validated on RX 580 (Vulkan), 4B @ 32k ctx: weights-only cold VRAM 1750 MB
(KV stays); weights+KV cold VRAM 726 MB (KV freed, ~1 GB reclaimed). KV
survives the round-trip: prompt cache reused after the cycle (prompt_n 4 vs
42), correct output.

Assisted-by: Claude
2026-11-21 19:39:46 +01:00
Lumpiasty 56a1010bbc server: on-demand VRAM sharing to time-share one GPU between models
Add release/restore of a model's GPU weight buffers (keeping a host shadow
and the KV cache) so several always-loaded llama-server processes can
time-share a single GPU without reloading or losing the prompt cache.

- llama-model: release_device_weights()/restore_device_weights() capture a
  compact host shadow (stable iteration order, view-skipping) and free then
  realloc the device weight buffers; weights_resident() query.
- llama-context: release_device()/restore_device() wrappers; decode() auto-
  restores; public C API llama_context_release_device/restore_device.
- server: LLAMA_SLEEP_VRAM_ONLY makes idle-sleep release only the VRAM weights
  (not a full unload/reload). A cross-process flock token in LLAMA_VRAM_ARENA
  enforces "resident iff holds token"; an inotify doorbell forces the holder
  to release on contention. The warden thread only touches the task queue, so
  releases run on the loop thread and never race a decode.

Validated on RX 580 (Vulkan): two models share 8GB, never both resident,
correct output under contention, KV cache preserved (no re-prefill).

Assisted-by: Claude
2026-11-21 19:32:30 +01:00
LumpiastyandClaude Opus 5.5 ab7e03ec6c llama: skip the routing-ids readback in sched_copy_experts when the batch uses every expert
Offloaded expert weights are uploaded per split, and to decide which experts to
upload the copy callback reads the routing ids back to the host. The ids are
produced on the same device we are about to upload to, so the readback forces a
full pipeline flush - 47 of them per eval on a 48-layer MoE.

Once the batch draws enough experts the readback stops telling us anything. At
2048 tokens x 10 experts over 256 experts every expert comes back used, so the
used set is all true and the copy is a single whole-tensor range anyway.

Skip the readback when ggml_nelements(ids) >= 4 * n_expert and mark every expert
used. Uploading an expert that no id selects cannot change the result, since
mul_mat_id only reads the rows the ids point at, so this stays exact. With the
MoE cache on, the all-used set still goes through moe_cache->copy_experts, which
copies the cached experts device-to-device from the host slot maps and uploads
the rest - still exact, and no readback is needed for it. Decode is unaffected:
n_ids there is the number of experts per token, far below the threshold. The
small-batch MoE cache slot path has its own readback and is not touched.

The st.ids/st.used caching is unchanged: a repeat call with the same ids tensor
reuses the all-true set, which is what the threshold would decide again.

Port of dbc449afd ("sched: skip the MoE routing-ids readback when the batch
uses every expert"), whose target block in ggml_backend_sched_compute_splits was
removed upstream (#29943) in favour of this copy callback.

Measured (original, before the port) on an RX 580 (Vulkan, Polaris) running
Laguna-S-2.1 118B IQ2_M with all experts host-resident: pp2048 +4.7% at depth
32k, +0.9% at 16k, tg32 +3%.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PZz44SLQvTXMyWGio6t9DZ
2026-11-21 19:32:13 +01:00
Xuan-Son Nguyen 19bc41dbb4 model: apply .scale tensors automatically + fix correctness (#32669)
* model: apply .scale tensors automatically

* rm build_moe_ffn "_s" overload

* use llama_prec_policy

* fix ci

* build_lora_mm_id: rm w_s
2026-11-21 16:30:01 +01:00
Xuan-Son NguyenandPascal c622f623fc mtmd: support kyutai/hibiki realtime full-duplex (#32491)
* mtmd: support hibiki realtime duplex

* fix and clean up

* fix long session quality

* fix ci

* tests : add the moshi fusion baseline for Metal

The Metal fusion check found 5 RMS_NORM+MUL fusions in the new moshi
arch and no baseline for them.

---------

Co-authored-by: Pascal <admin@serveurperso.com>
2026-11-21 14:25:19 +01:00
Daniel Han 690b87d881 model : inkling2_loop share the loop slots after the scale pass (#32704)
The loop slots copied the physical layers at the end of load_arch_tensors,
before the generic pass attaches the optional .scale tensors, so the later
passes through the core ran NVFP4 experts without their per-expert scales.
Add an optional post_arch_tensors hook that runs after the scale pass and
share the slots there.
2026-11-21 12:10:38 +01:00
Sigbjørn Skjæret 838d942438 model : additional shortcut checks for OOB in hrm_text (#32720)
[no ci]
2026-11-21 12:10:04 +01:00
Georgi Gerganov e412c6f468 llama : remove fused lightning indexer gating (#32650)
The fused lightning indexer is always used now, so drop the cparams flags, the auto probe and the non-fused fallbacks.

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
2026-11-21 10:26:55 +02:00
Daniel Kuts 46eb1742db llama-grammar: Fix integer overflow (#32551)
* llama-grammar: Fix integer overflow

* Check incorrect values at parsing and bail out nicely

* fix style
2026-11-21 07:00:30 +01:00
Aman Gupta 6792507b7f llama : score only the k-pools of the ubatch sequences in QSA (#32519) 2026-11-21 13:22:36 +08:00
Pedro Cuenca 29b52d398b model : Inkling 2 Mini (#32694)
* Conversion

* Separate model_type

* Add gate_rank to allow building from metadata while testing

* Base model

* Loop graph

* Support the new is_varloop attribute

* Adopt naming scheme from https://github.com/ggml-org/llama.cpp/pull/32588

* Rename to varying
2026-11-20 23:01:07 +01:00
Meet Amin 184862801a gemma4: Optimized prompt processing for multi-seq queries (#32342) 2026-11-20 23:41:02 +08:00
Daniel HanandSigbjørn Skjæret a1f4222608 model : load ffn_gate_up_exps scales (#32657)
* model: load `ffn_gate_up_exps` scales

The generic scale pass loads per-expert scales for ffn_gate_exps,
ffn_up_exps and ffn_down_exps but not for the merged ffn_gate_up_exps.
Load ~ffn_gate_up_exps_s there~ and expose it as ffn_up_exps_s ~when the layer
has no separate up tensor,~ since build_moe_ffn applies up_exps_s to the
merged gate_up output. Inkling 2 now passes the expert scales to
build_moe_ffn like the other MoE models.

* Update src/llama-model.cpp

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

* Update src/llama-model.h

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>

---------

Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
2026-11-20 13:11:28 +01:00
Aman Gupta 4783b1145c llama : move the indexer key rotation choice into the model classes (#32653)
* llama : move the indexer key rotation choice into the model classes

The KV cache forced a Hadamard rotation for the lightning indexer keys
based on a list of architectures. The virtual llama_model::indexer_k_rot()
now returns true for the same models. DeepSeek4.1 inherits from DeepSeek4
but was not in the list, so it overrides it back to false.

* llama : rename indexer_k_rot() to lid_k_hdmd()
2026-11-20 18:49:16 +08:00
Aman Gupta 652d559edf llama : move the graph node budget into the model classes (#32648)
graph_max_nodes() picked the nodes-per-tensor factor from lists of
architectures. Each model class now gives it through a new virtual,
llama_model::graph_nodes_per_tensor(), which defaults to 8. The
architectures that need 32 or 64 override it in models.h, and DFlash
keeps its hparams-based rule in dflash.cpp.

LoRA adapter nodes are now added for every architecture, not only for
the ones on the default factor.
2026-11-20 17:19:14 +08:00
Aman Gupta 4c4a0f0a42 llama : reserve graph nodes for the MoE cache (#32636)
The MoE cache adds a slot lookup to every MoE layer, and with
--moe-cache-cpu-share a second expert path, but graph_max_nodes() did
not account for it. Models on the default 8*n_tensors budget with few
tensors per layer (cohere2moe: 443 tensors, 49 layers) overflow the
graph context when the cache is combined with a quantized KV cache
(K/V rotation), aborting in ggml_new_tensor with GGML_ASSERT(obj_new).

Fixes #32584

Assisted-by: Claude Opus 5.5
2026-11-20 16:13:00 +08:00
nicholas-mainardi-domyn 55d323507d model: add Domyn Edge 4B Model (#32559)
* feat: add domyn-edge model

* Generate tokenizer hash for conversion + fix tests

* Address changes according to code-review skill

* catch ValueError in convert_hf_to_gguf_update.py

* renegerate fusion test baseline + fix model naming style
2026-11-20 09:51:41 +02:00
Imdadullah Mohammed 50518b64c6 model : hrm : do not compute slot count as uint32_t (#32601)
load_arch_hparams multiplies n_hrm_layers_per_stack, n_hrm_h_cycles and
(n_hrm_l_cycles + 1) into a uint32_t, then asserts the product equals
block_count.

Any factors whose true product is block_count + k*2^32 still pass the
assert; 1073741825 * 2 * 2 = 2^32 + 4 compares equal to block_count = 4.

load_arch_tensors iterates the same three values in int and writes past
the end of the block_count-sized layers vector.
2026-11-19 23:21:03 +01:00
Piotr Wilkin (ilintar) 8af68d486a llama: fixes for Inkling 2 attention mask (#32429)
* llama : give Inkling2 its own relative-position attention input

The Inkling2 graph built its flash attention mask from the hybrid iswa
input and passed it to build_attn through a new kq_mask argument. That
argument let a caller pair its own dense mask with the kv_rows of
llm_graph_input_attn_kv_iswa, whose mask follows the per-sequence cell
lists whenever kv_rows is enabled.

Add llm_graph_input_attn_rel_pos instead. It owns dense KQ masks and the
query-to-cell distances, fills both from the same cells, and its
build_attn overload adds the per-head bias to its own mask and never
passes kv_rows. The iswa build_attn is back to its previous signature;
its body moves to build_attn_iswa_impl, which both overloads call.

Inkling2 now folds the per-query sink into the bias with and without
flash attention, so build_attn_mha no longer needs the per-query sink
path.

Assisted-By: Claude Opus 5.5 <noreply@anthropic.com>

* inkling2 : make the expert weight scale and final logit softcap optional

Load both keys as optional and skip them when they are zero, as other
models do, instead of rejecting zero at load.

Assisted-By: Claude Opus 5.5 <noreply@anthropic.com>

* llama : build kv_rows masks for the hybrid iswa memory

llm_graph_input_mem_hybrid_iswa checked its masks with
can_reuse_attn_mask, which expects kv_rows whenever they are enabled,
but built them dense. With a unified cache and flash attention the
check always failed, and hybrid iswa models (LFM2, LFM3, Inkling2)
never reused a graph. Build the masks with build_attn_inp_attn_mask,
like the plain iswa input.

Assisted-By: Claude Opus 5.5 <noreply@anthropic.com>

* llama : make the relative-position input a hybrid iswa input

Address review: llm_graph_input_mem_hybrid_iswa_rel derives from
llm_graph_input_mem_hybrid_iswa and owns the attention and recurrent
inputs, so Inkling2 builds one memory input instead of two. Its
attention input is built with dense masks, can_reuse checks them as
dense, and its build_attn asserts that the mask has no kv_rows (TODO:
a kv_rows variant). build_attn_iswa_impl is now a static helper.

Assisted-By: Claude Opus 5.5 <noreply@anthropic.com>

* llama : compose the hybrid iswa rel input, format inkling2

Address review: llm_graph_input_mem_hybrid_iswa_rel owns its attention
and recurrent inputs like the other composite inputs instead of deriving
from llm_graph_input_mem_hybrid_iswa.

Format src/models/inkling2.cpp like the other model sources: grouped
and aligned key and tensor loading, descriptive per-layer names, blank
lines and section comments between the steps, and cb() names for the
intermediate tensors. The graph is unchanged.

Assisted-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-19 14:36:30 +01:00
Georgi Gerganov 01592dc58b llama : remove redundant fields from llm_graph_params (#32596)
llm_graph_params no longer carries hparams, prec_policy and hdmd, since the model reference already provides them. llm_graph_context sources them from the model and keeps prec_policy as a reference.

Assisted-by: dsh:llama.cpp/ggml-org/DeepSeek-V4.1-Flash-1017-GGUF (Mac Studio, M5 Ultra, 256GB)
2026-11-19 13:31:50 +02:00
saiph 51a77c4f24 qwen3vl : set t_layer_inp for EAGLE-3 (#32589)
Assisted-by: Grok Bot
2026-11-19 10:13:02 +01:00
Adrien Gallouët 95b09b62be llama : remove PATH_MAX (#32557)
Signed-off-by: Adrien Gallouët <angt@huggingface.co>
2026-11-19 08:57:57 +01:00
Adrien Gallouët 39039aa27f llama : fall back to buffered I/O on EIO (#32545)
* llama : fall back to buffered I/O on EIO

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

* Preserve bytes already read, resumes at the current file offset

Signed-off-by: Adrien Gallouët <angt@huggingface.co>

---------

Signed-off-by: Adrien Gallouët <angt@huggingface.co>
2026-11-19 07:50:33 +01:00
mahounairoandAlexey Kopytko 0a76757469 llama : fix sticky NaN in recurrent state (#31913)
Co-authored-by: Alexey Kopytko <alexey@kopytko.com>
2026-11-19 08:08:12 +09:00
parthod0x 78e05298f0 model : limite: reject mudd configs that index past mudd_tap_idx on load (#32562)
n_mudd_taps comes from limite.mudd.tap_count, a scalar read separately
from the length-checked tap_indices array. mudd_tap_idx is a fixed
std::array<int32_t, 512> indexed at [i*n_taps] in both the tensor load
and the graph build, so a crafted GGUF with n_layer * n_mudd_taps > 512
reads past the array.

Reject that at hparams load; the one check covers both index sites.
2026-11-18 23:09:55 +01:00
Tarek Dakhran bafa0420c1 server : enable packing for d1-2B and d1-Omni-2B models (#32553)
* llama-bench : size the batch to the prompt for models without memory

Assisted-by: Claude Opus 5.5

* server : read up to 16 lfm3-d1 questions in one pass

Assisted-by: Claude Opus 5.5
2026-11-18 20:15:43 +01:00
varunandvarun 98d67c8d53 model:add Zaya1 8B (Zyphra) (#32071)
* model:add Zaya1 8B (Zyphra)
- Added the support for Zaya.
- Fixed the SSM_CONV case, so the model runs on cuda without abort.

 Please enter the commit message for your changes. Lines starting

* Made the following changes based on the bot reviewer feedback:
1) Added runtime asserts when loading wrong KV head size, unsupported conv kernel size.
2) Removed the static cast usage.
3) Fixed double load of tokenizer.
4) Updated res_scale_in to be input tensor.

* - Moved the res_scale_in back to layer 0 to fix the cuda bug

* i)  Removed the EOS addition.Team agreed that updating Tokenizer is the right move.
ii) Fixed formatting

* Fixed the comment to make it clear

---------

Co-authored-by: varun <varun_subramanian@zyphra.com>
2026-11-18 15:46:49 +01:00
Gaurav Garg bba8408821 Add LLAMA_MOE_RANDOM_ROUTING=1 for benchmark-only random expert selection (#29539)
* Add LLAMA_MOE_RANDOM_ROUTING=1 for benchmark-only random expert selection

* Handle expert caching
2026-11-18 17:04:40 +05:30
Daniel Bevenius 61f182fcde unicode : add continuation check lambda (#31070)
This commit adds a lambda to check if a specified byte is not a valid
continuation byte in a multibyte UTF8 sequence.

The motivation for this is to improve readability and avoid some code
duplication.
2026-11-18 11:44:28 +01:00
Daniel Bevenius 7144124a5b vocab : extract bpe_rank loading into function (#30962)
* vocab : extract bpe_rank loading into function

This commit extracts the BPE rank loading logic into a separate function
named `load_bpe_ranks`.

The motivation for this is that the came code is currently used by both
gpt2/hybriddna/whitespace tokenizer models and gemma4 and this avoids
the duplication of code to improve readability and maintainability.

* change signature of load_bpe_ranks
2026-11-18 11:33:12 +01:00
Xuan-Son Nguyen 3574df84b7 mtmd: support tiny-voca (#32427)
* mtmd: support tiny-voca

* improve G2P
2026-11-18 00:04:15 +01:00
Xuan-Son Nguyen 307e3f1551 mtmd: add kokoro tts support (#32413)
* jinja: impl scope-chain key lookup

* add test for call_statement lambda

* chat: add tts_dict for kokoro G2P template

* more testing

* value_float_t range check

* no json-in-meta, add llama_model_meta_arr_str/_n

* mtmd: add kokoro tts support

* detailed guide about G2P and text norm port

* refine G2P

* clean up

* migrate batch api

* nits
2026-11-17 22:25:44 +01:00
Aman Gupta ad9d9e8f50 qwen4exp : fix graph reallocation from the n-gram table scale (#32483)
* qwen4exp : apply the n-gram table scale with ggml_scale

* read the scale once after load

* remove comment
2026-11-18 00:35:54 +08:00
08a9b35893 llama : keep the k-pool layout prefix on sequence edits (#32137)
* llama : keep the k-pool layout prefix on sequence edits

Assisted-by: Claude Code

* llama : address review of the k-pool layout update

* Append and edit share one branch, using stale > sq.pos_min.
* LLAMA_KPOOL_VERIFY=1 checks every update against a full rebuild. The scan moved into kpool_scan() so the check can reuse it.
* The test turns the check on for its rollback context, skips -sm tensor, and drops kimi-k31.
* Two comments that were no longer true are fixed.

Assisted-by: Claude Code

* llama : check the k-pool layout update in debug builds instead of LLAMA_KPOOL_VERIFY

* llama : remove the k-pool layout verification

* llama : inline the k-pool scan again

* remove dangling new line

* do not search twice

---------

Co-authored-by: Aman Gupta <amangupta052@gmail.com>
Co-authored-by: Alexey Kopytko <alexey@kopytko.com>
2026-11-17 22:37:25 +08:00
Hiroyuki Ikegami acad5e405c vocab : speed up PLaMo-2/3 tokenizer table build (#32249)
Assisted-by: Claude Code (Claude Opus 5.5)
2026-11-17 15:36:15 +01:00
Aman Gupta d3064fddce convert : fix FP8 qwen4exp checkpoints (n-gram table scale, block-scaled FP8) (#32476)
* convert : fix FP8 n-gram table export for qwen4exp

* convert : dequantize block-scaled FP8 weight_scale_inv in ModelOpt checkpoints
2026-11-17 22:25:01 +08:00
Xuan-Son Nguyen 8d1599fca1 chat: add tts_dict for kokoro G2P template (#32397)
* chat: add tts_dict for kokoro G2P template

* more testing

* value_float_t range check

* no json-in-meta, add llama_model_meta_arr_str/_n
2026-11-17 12:52:50 +01:00
cherry77-cloud 4df3f2959c qwen3next : apply output projection weight scale (#32393)
Assisted-by: Codex
2026-11-17 11:32:30 +01:00
Aldehir Rojas 4ddf2b9470 chat : implement inkling parser (#32381) 2026-11-16 19:48:24 +01:00
Niklas Wenzel 62a6d3a302 llama : migrate to llama_batch_ext (#32356)
* llama : migrate to llama_batch_ext

* mtmd : migrate zonos3-mini code

Added in https://github.com/ggml-org/llama.cpp/pull/32020
2026-11-16 15:16:12 +01:00
Titaniumtown 21fa9c043f kv-cells: avoid an O(n_cells) scan per sequence removal (#32380) 2026-11-16 10:52:09 +01:00
Seungjun Bae 18b74fa41d model : read qwen3vl output head size from the tensor (#32349)
Allows Qwen3-ForcedAligner, whose output head (5000 timestamp classes) is smaller than the vocab.
2026-11-15 19:45:17 +01:00
Xuan-Son Nguyen 7ae3c6b231 mtmd: add zonos3-mini TTS model (#32020)
* zonos3 init

* add audio gen pipeline

* fix string_format

* add default voice as gguf metadata

* better ref

* support voice clone

* add delivery params

* optimize speed

* nits

* nits 2

* add text normalization

* nits
2026-11-15 19:44:44 +01:00
Xuan-Son Nguyen a68799c609 model: support MiniMax-M3.1 (#32301) 2026-11-15 17:39:59 +01:00
Anton SokolchenkoandJames Pickering de32b268c4 Support DFLASH2 for Recursive Ripple-65B-Preview (#31904)
* fixed support for DFlash2 for Ripple 65B

* spec: Add NoPE handling to DFlash

(cherry picked from commit 89ea8e3e77e3359d17902a2db136022d5d419d88)

* spec: Fix NoPE not being picked up by DFlash

(cherry picked from commit 43393b95fda2ce7e7d243e2650bf7b1cadf1c732)

---------

Co-authored-by: James Pickering <james_pic@hotmail.com>
2026-11-15 19:36:08 +08:00
Tarek Dakhran e574a17b06 model : metal : workaround fusion for LFM3 (#32296)
* lfm3 : put the normed tensor first in residual adds to enable norm fusion

Assisted-by: Claude Opus 5.5

* tests : update Metal fusion baseline for lfm3

Assisted-by: Claude Opus 5.5
2026-11-15 01:05:20 +01:00
Piotr Wilkin (ilintar)andXuan Son Nguyen 5c008e0008 feat: add Inkling 2 support (text + vision) (#32217)
* convert : add Inkling2 text GGUF conversion

Assisted-by: Codex

* llama : implement Inkling2 text inference

Assisted-by: Codex

* tests : cover Inkling2 text support

Assisted-by: Codex

* conversion : dequantize packed Inkling2 experts in row chunks

Decode per-layer NVFP4 and MXFP8 schemes, validate companions, and use stored scale shapes with a warning for the declared MXFP8 [32,32] grid. Reorder packed w13 rows and their scale rows together. Keep normal GGUF outtype selection and 3D expert tensors.

Stream 128 rows per expert through LazyChunkedTensor. Real w13 (384x4608x4608) needs a 2.25 MiB F32 chunk instead of a 30.375 GiB expanded layer. Real-shaped sparse probes peak below 5.8 MiB of Python/NumPy chunk allocations, plus 5.2 MiB of callback metadata: budget about 16 MiB per projection including mapped pages, excluding runtime, other tensors and OS cache. Full-model chunk metadata is roughly 370 MiB.

Packed synthetic GGUF weights match the independent pinned vLLM decode bit-for-bit. CPU prefill/decode logits remain within 3.1e-6 absolute error, with 100% argmax agreement. Float fixtures, original-name conversion and CPU architecture checks pass. Real checkpoint conversion remains untested.

Assisted-by: Codex

* convert : add Inkling2 vision mmproj support

Assisted-by: Codex

* mtmd : implement Inkling2 still-image encoder

Assisted-by: Codex

* inkling2 : scope image loading

Assisted-by: Codex

* inkling2 : enable flash attention with folded bias and sinks

Assisted-by: Codex

* cuda : route per-head flash attention masks through vec

Assisted-by: Codex

* mtmd : support Inkling2 temporal frame means

Assisted-by: Codex

* style : restore shared image comment

Assisted-by: Codex

* tests : enlarge Inkling2 context fixture

Assisted-by: Codex

* inkling2 : load per-layer expert feed-forward widths

Assisted-by: Codex

* tests : add Inkling2 Metal fusion baselines

Assisted-by: Codex

* inkling2 : group and align registry entries

Assisted-by: Fable

* convert : regenerate the Inkling2 tokenizer hash with the update script

Assisted-by: Fable

* conversion : revert the print_registered_models change

Assisted-by: Fable

* conversion : sort the Inkling2 model map entries

Assisted-by: Fable

* convert : map Inkling2 vision checkpoint names in the tensor table

Assisted-by: Fable

* convert : map Inkling2 text checkpoint names in the tensor table

Assisted-by: Fable

* tests : document the Inkling2 packed expert conversion test

Assisted-by: Fable

* tests : cover Inkling2 checkpoint and transformers tensor names

Assisted-by: Fable

* llama : place Inkling2 fields and list entries by file convention

Assisted-by: Fable

* gguf-py : align Inkling2 tensor mapping comments

Assisted-by: Fable

* mtmd : place Inkling2 entries by file convention

Assisted-by: Fable

* tests : place Inkling2 cases after the existing ones

Assisted-by: Fable

* cuda : index per-head flash attention masks in the tile and mma kernels

Assisted-by: Claude
Claude-Session: https://claude.ai/code/session_01GbYSza1yxgmTz8wdr6uV8Q

* inkling2 : project the relative bias before the gather, drop arch branches

Assisted-by: Claude
Claude-Session: https://claude.ai/code/session_01GbYSza1yxgmTz8wdr6uV8Q

* convert : require Inkling2 hparams, add inkling2 pre-tokenizer

Assisted-by: Claude
Claude-Session: https://claude.ai/code/session_01GbYSza1yxgmTz8wdr6uV8Q

* mtmd : cap Inkling2 image size, drop video frame averaging

Assisted-by: Claude
Claude-Session: https://claude.ai/code/session_01GbYSza1yxgmTz8wdr6uV8Q

* inkling2 : fix CI build, Meta device assert and Windows conversion test

Assisted-by: Claude

* tests : use the default sliding window for the Inkling2 test model

Assisted-by: Claude

* simplify mtmd part

* give _v arch for vision

* remove alpha channel handling, it should be a follow up

* explicit rule about audio + vision models

* add audio support

* rm audio

* rm gguf mtmd rope base

* address review comment

* nits

* do not normalize embd from mtmd

---------

Co-authored-by: Xuan Son Nguyen <son@huggingface.co>
2026-11-15 00:51:31 +01:00