Adds a new `mtp` boolean to `llama_model_params`. When set to false (default):
1. The loader skips loading MTP-specific tensors (NextN layers) using `TENSOR_SKIP`.
2. The KV cache size calculation excludes the MTP layer (`n_layer_kv_from_start`).
This reduces VRAM usage and load time for users running GLM-4.5/4.6 in standard generation mode.
commit 912ed2cd9339d1b2875d98744ca5b51fa62e581e
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Dec 7 23:00:29 2025 -0300
speculative (feat): implement recursive MTP drafting for GLM-4.5
commit bdf72d9552e3da64ffc85f175664713388752914
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Dec 6 16:10:16 2025 -0300
sampling (feat): optimize speculative drafting with fast-path selection
commit a91980a8f3475a6bbac0a64d8be06dd4b613020e
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Dec 6 15:18:19 2025 -0300
mtp (chore): clean old code
commit 6de0ecf55db8567db4faa99b0152b72c9e854548
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Dec 6 14:40:13 2025 -0300
mtp (feat): add mtp arg
commit ea77394183b8e6c368af969b8274039a54b11486
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Dec 6 13:47:54 2025 -0300
mtp-graph (fix): move llama_get_logits_ith outside the loop
commit 15dff208958fb66802f20ec53ce5fcaff133edb7
Merge: 171346c74 cae85fe53
Author: samuel <samueloliveira32df@gmail.com>
Date: Thu Oct 16 13:44:41 2025 -0300
Merge branch 'glm4-mtp-batch' of https://github.com/SamuelOliveirads/llama.cpp into glm4-mtp-graph-cache
commit cae85fe531
Author: samuel <samueloliveira32df@gmail.com>
Date: Thu Oct 16 13:42:31 2025 -0300
mtp-batch(fix): avoid logits for mtp kv cache operations
commit 171346c742c310bbcfbd786b61250638ccf8b44d
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Oct 12 16:33:01 2025 -0300
mtp-graph(feat): Reactivate graph reuse only for main model path
commit 0127c6beeb
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Oct 11 22:20:54 2025 -0300
mtp-batch(chore): Remove final MTP debug logs and dead code
commit 4bcc9e261e
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Oct 11 18:51:22 2025 -0300
mtp-batch(fix): Correctly advance cache head and add MTP documentation
commit b4cbe030ac
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Oct 11 18:37:40 2025 -0300
mtp-batch(chore): Fix logit flags for speculative sampling and remove debug logs
commit a99709d0c1
Author: samuel <samueloliveira32df@gmail.com>
Date: Fri Oct 10 17:24:34 2025 -0300
mtp-batch(refactor): Extract decode context and MTP input logic into helper methods
commit 913af8f48d
Author: samuel <samueloliveira32df@gmail.com>
Date: Fri Oct 10 16:44:28 2025 -0300
mtp-batch(refactor): Replace MTP boolean flags with an explicit operation enum
commit 6f74ba3807
Author: samuel <samueloliveira32df@gmail.com>
Date: Thu Oct 9 22:27:18 2025 -0300
mtp-batch (fix): prevent mtp draft from polluting the cache
commit 5e1d719bef
Author: samuel <samueloliveira32df@gmail.com>
Date: Thu Oct 9 15:21:23 2025 -0300
mtp-batch (feat): Create and manage sinfo for MTP
commit febd8235d2
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Oct 5 14:43:40 2025 -0300
mtp-batch (wip): fix how to warmup kv cache for MTP
commit 67c6c069e0
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Sep 27 19:42:32 2025 -0300
mtp-batch (wip): Isolate MTP graph to prevent host embedding buffer corruption
commit 75dc25e6fe
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Sep 27 17:17:00 2025 -0300
mtp-batch (wip): organize batch for mtp cache
commit 3da7e7f330
Author: samuel <samueloliveira32df@gmail.com>
Date: Tue Sep 23 22:45:11 2025 -0300
mtp-batch (fix): warm mtp cache for small batch size
commit df64508b93
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Sep 21 21:55:41 2025 -0300
mtp-batch (wip): merge glm graphs
commit 042eb8a829
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Sep 21 21:29:00 2025 -0300
mtp-batch (wip): merge mtp and model graph
commit 1318b2de82
Author: samuel <samueloliveira32df@gmail.com>
Date: Sun Sep 14 10:22:59 2025 -0300
mtp-batch (wip): move mtp execution to batch format
commit c6237c71ff
Merge: 9fab53e438742ce0e3
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Sat Sep 13 02:57:01 2025 -0400
Merge pull request #1 from SamuelOliveirads/glm4-moe-mtp
feat: implemented sampling for MTP
commit 8742ce0e39
Author: samuel <samueloliveira32df@gmail.com>
Date: Sat Sep 6 00:21:18 2025 -0300
feat: apply logits + greedy sampler
commit 5a5bce8577
Author: samuel <samueloliveira32df@gmail.com>
Date: Wed Sep 3 17:56:14 2025 -0300
fix: add sample acceptance
commit 07670a22c6
Author: samuel <samueloliveira32df@gmail.com>
Date: Wed Sep 3 13:25:21 2025 -0300
feat: implemented sampling for MTP
commit 9fab53e438
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Tue Sep 2 17:14:09 2025 -0400
fixed mtp kv cache update step in cases where prompt size > n_batch and n_ubatch
commit 98bc0c6bf2
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Tue Aug 26 01:26:51 2025 -0400
replace standard sampler with greedy sampler for mtp draft
commit 471e026327
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Tue Aug 19 23:10:56 2025 -0400
fixed vram leak
commit d72f9d5691
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Tue Aug 19 01:50:34 2025 -0400
kludge-y kv cache management of mtp layer
commit 382135aa36
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Sun Aug 17 21:54:45 2025 -0400
fixed mtp kv cache update sequencing after prompt processing
commit 6870f9790c
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Sun Aug 17 04:59:36 2025 -0400
added proper KV cache management for MTP layers and slightly refactored
commit 6e9bafc7a7
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Fri Aug 15 23:13:56 2025 -0400
failed attempt to implement MTP; outputs tokens but KV cache management is unreasonable
commit cf0f7c0448
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Wed Aug 13 02:21:17 2025 -0400
broad thrust of the mtp implementation
commit 03231da69e
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Tue Aug 12 01:03:59 2025 -0400
add model member function to build mtp graph, to be called from speculative.cpp
commit 1f477b3755
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Mon Aug 11 20:54:45 2025 -0400
make nextn weights loadable without a crash
commit e434f87cc7
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Mon Aug 11 01:21:47 2025 -0400
some work towards building mtp layer graph
commit db60623e79
Author: Aaron Lee <lee.aaron.65@gmail.com>
Date: Sun Aug 10 23:52:54 2025 -0400
added getter for nextn layer count and server slot has_mtp property
* llama: automatically fit args to free memory
llama-fit-params tool
* fix CI
* hints for bug reports, ensure no reallocation
* fix segfault with Vulkan
* add llama-fit-params to CI
* fix CI
* fix CI
* fix CI
* minor adjustments
* fix assignment of 1 dense layer
* fix logger not being reset on model load failure
* remove --n-gpu-layer hint on model load failure
* fix llama-fit-params verbosity
* fix edge case
* fix typo [no ci]
* kv-cache : pad the size of the small SWA cache for performance
* context : pad the total context to 256
* cont : future-proof the swa pad
* server : adjust test params to new logic
* server : support unified context across slots
* cont : fix speculative decoding initialization
* context : fix n_ctx_per_seq computation
* server : purge slots one by one
* tests : add unified cache server tests
* llama : update per-seq context computation
* test-thread-safety : handle tiny training context of the input model
* server : fix server_tokens clear()
* server : use 4 slots + unified KV by default
* llama : add note about context size queries
* cont : update todos [no ci]
* context : do not cap the size of the context
* tests : adjust parameters to be CI friendlier
* context : add warning
commit 5fb5e24811 (llama : minor
sampling refactor (2) (#9386)) moved the llama_sampler_accept call
into llama_sampler_sample, but the sampling sample usage in llama.h
was forgotten to be updated accordingly.
* implement --no-host to disable host buffer
* fix equal_mparams
* move no-host enumeration order together with other model params
---------
Co-authored-by: slaren <slarengh@gmail.com>
* initial commit for branch 3
* generalize `swa_checkpoint` to `ctx_checkpoint`
this extends `llama-server`'s SWA checkpointing logic to include
hybrid/recurrent models such as Jamba, Granite
* oops
* disable debug prints
* keep backwards compat with `--swa-checkpoints`
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* update prompt re-processing message
* fix off-by-one error per GG
* keep `seq_rm` log per GG
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* server : fix checkpoint logic to support recurrent caches
* server : cleanup and fixes
---------
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* feat: Add python-side constants and conversion for adapter.lora.invocation_string
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add c++ side constants for adapter.lora.invocation_string
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Parse invocation string for adapters from GGUF
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(python): Update conversion to alora_invocation_tokens
This is the preferred method in PEFT which is the source of ground truth
https://github.com/huggingface/peft/pull/2609/files#diff-13380145401d203d5935c5189dd09879f990b81aa63e8e3aaff8ce9110333f0e
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix(cpp): Update to alora_invocation_tokens on c++ side
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Add C APIs to get alora invocation token array from lora
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Initial implementation of alora cache logic in server
This does not yet do the part to identify the invocation tokens and only
apply the lora adapter afterwards, but it does seem to produce correct
results if the invocation tokens are the beginning of the uncached input.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Identify alora invocation sequences
This currently limits to a single enabled alora per slot. Multiple aloras
with different invocation sequences would be possible, but it would require
a more complex integration of the adapter toggling and is not really a well
studied case for alora since it's unclear if one alora can reuse cache from
previous prefill computed with a different alora.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Only reuse cache for tokens before the alora invocation start
This is a bit of an edge case, but theoretically a user could try the same
query with the alora disabled (just using the base model), then retry with
the alora. The cached tokens from the first pass should be invalid.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* feat: Handle un-cached tokens that come before the alora activation
The solution is to only fill up to the token before the invocation start in
the batch if there are any tokens to be prefilled between those pulled from
cache and the invocation start. When this is detected, the alora is
temporarily disabled with a scale of 0.0, then immediately re-enabled after
it has been initialized for the internal graph. Since the batch does not
complete the prompt tokens, the remaining prompt tokens are handled in the
next task, pulling all of the non-alora tokens from cache and proceeding
with prefill for the alora tokens.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Use || instead of 'or'
Too much python 🤦
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Fix off-by-one for limiting cached tokens to before alora start
This was the cause of the inconsistent results from the dummy test script
with and without the turn that runs the prompt without the adapter before
running it with the adapter.
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Support backwards-compatibility for "invocation_string" in adapter_config.json
While this has been replaced in the PEFT PR in favor of
alora_invocation_tokens, the existing adapters in the ibm-granite org on HF
use "invocation_string," so this will enable backwards compatibility and
enable testing now (before PEFT PR changes have percolated everywhere).
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
* fix: Remove duplicate logging
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* feat: Report alora_invocation_string and alora_invocation_tokens from /lora-adapters
Branch: gabe-l-hart/alora-support
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
---------
Signed-off-by: Gabe Goodhart <ghart@us.ibm.com>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* sampling : optimize sorting using bucket sort in more places
ggml-ci
* sampling : do not sort in dist sampler
ggml-ci
* sampling : avoid heap allocations for sort buffers
ggml-ci
* common : add option to sort sampling candidates by probability
ggml-ci
* sampling : revert the change for preserving sort buffers
* sampling : use std::copy instead of memcpy
* sampling : clarify purpose of partial sort helpers
ggml-ci
* cont : remove wrong comment [no ci]
* common : update comment
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
---------
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* server : add SWA checkpoints
ggml-ci
* cont : server clean-up
* server : handle state restore fails
* llama : add extended llama_state_seq_ API
* server : do not make checkpoints if --swa-full
ggml-ci
* llama : remove flags value for NONE
* server : configure number of SWA checkpoints with CLI arg
ggml-ci
* args : fix scope of new argument
* examples/finetune -opt SGD (stochastic gradient descent) memory opt
add unit tested GGML_OPT_OPTIMIZER_SGD to ggml - avoids allocating
m, v tensors.
support finetune.cpp arg -opt SGD (or sgd). (default adamw as before)
llama 3.2-1b-F32 result: observed 11gb gpu ram (41 sec/epoch)
when using SGD instead of 19gb (55 sec/epoch) using adamw.
(wikipedia 100 lines finetune)
(
using the same GPU memory, adamw can only do before OOM 512
batch/context, reaching:
train: [███████▉] data=0000140/0000140 loss=0.02575±0.00099 acc=99.52±0.03% t=00:00:47 ETA=00:00:00
val: [███████▉] data=0000008/0000008 loss=4.76565±0.28810 acc=41.46±0.77% t=00:00:00 ETA=00:00:00
SGD is superior, though it converges slower, with max before OOM 1728
batch/context (esp see the better validation perf):
train: [███████▉] data=0000039/0000039 loss=0.00371±0.00010 acc=99.96±0.01% t=00:00:41 ETA=00:00:00
val: [███████▉] data=0000003/0000003 loss=5.11406±0.76034 acc=48.01±0.69% t=00:00:01 ETA=00:00:00
)
note: when finetuning long enough (or w/ enough -lr),
validation accuracy *eventually* drops ('catastrophic forgetting')
-lr-half (halflife) option useful for SGD to avoid oscillation or
super slow underdamped learning (makes setting -lr more forgiving).
terminal -lr for now is set by lr-halvings i.e. if you want at most
1/8 the inital -lr you set -lr-halvings 3.
note: objective loss not directly comparable between adamw, sgd? -
check perplexity or accuracy or consider relative improvements
for convergence
new finetune args -wd 1e-9 to enable weight decay in sgd or adamw,
and max -epochs N (default 2 as before)
cache (1 - wd*alpha) in 'adamw' opt struct -
no noticeable perf benefit, disabled (still done
for new SGD though)
since opt. memory is pre-allocated, the ggml_opt_get_optimizer_params
would probably be able to change between SGD and AdamW with each epoch
but would need to use adamw for the first (unconfirmed - no cmdline arg
to set such a policy yet)
test-opt checks adamw as before and now sgd (except for a few disabled
tests for sgd only; probably just needs logging values and adding
alternate reference values); tolerance on the 'regression'
test is broader for sgd (so we don't need many more epochs)
* Vulkan: Implement GGML_OP_OPT_STEP_SGD
* tests: Fix OPT_STEP_SGD test-backend-ops
* SGD op param store weight-decay and not 1-alpha*wd
* minor + cosmetic changes
* fix vulkan sgd
* try CI fix
---------
Co-authored-by: 0cc4m <picard12@live.de>
Co-authored-by: Johannes Gäßler <johannesg@5d6.de>
* Add support for Llada-8b: diffusion model
* Add README
* Fix README and convert_hf_to_gguf
* convert_hf_to_gguf.py: address review comments
* Make everything in a single example
* Remove model-specific sampling
* Remove unused argmax
* Remove braced initializers, improve README.md a bit
* Add diffusion specific gguf params in set_vocab, remove setting rope_theta and rms_norm_eps
* Remove adding the mask token
* Move add_add_bos_token to set_vocab
* use add_bool in gguf_writer.py
* Support diffusion models: Add Dream 7B
* Move diffusion to examples
* Move stuff to examples. Add patch to not use kv-cache
* Address review comments
* Make sampling fast
* llama: remove diffusion functions
* Add basic timings + cleanup
* More cleanup
* Review comments: better formating, use LOG instead std::cerr, re-use batch, use ubatch instead of max_length
* fixup!
* Review: move everything to diffusion-cli for now
Add LLAMA_API to fix the run-time error with llama-cpp-python in Windows env:
attributeError: function 'llama_kv_self_seq_div' not found.
Did you mean: 'llama_kv_self_seq_add'?
Although llama_kv_self_seq_div() has been marked deprecated but
it is necessary to export it to make llama-cpp-python happy.
Observed software version:
OS: windows
compiler: MSVC
llama-cpp-python: tag: v0.3.12-cu124
llama.cpp: tag: b5833
Signed-off-by: Min-Hua Chen <minhuadotchen@gmail.com>
Co-authored-by: Min-Hua Chen <minhua.chen@neuchips.ai>
* Add PLaMo-2 model using hybrid memory module
* Fix z shape
* Add cmath to include from llama-vocab.h
* Explicitly dequantize normalization weights before RoPE apply
* Revert unnecessary cast because the problem can be solved by excluding attn_k, attn_q when quantizing
* Use ATTN_K/Q_NORM for k,q weights to prevent quantization
* Remove SSM_BCDT that is not used from anywhere
* Do not duplicate embedding weights for output.weight
* Fix tokenizer encoding problem for multibyte strings
* Apply suggestion from @CISC
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update src/llama-model.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Use LLM_FFN_SWIGLU instead of splitting ffn_gate and ffn_up
* Remove unnecessary part for Grouped Query Attention
* Fix how to load special token id to gguf
* Remove unused tensor mapping
* Update src/llama-model.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Remove llama_vocab_plamo2 class and replace it with llm_tokenizer_plamo2_session to follow the other tokenizer implementations
* Update src/llama-vocab.cpp
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update src/llama-model.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update src/llama-model.cpp
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Update convert_hf_to_gguf.py
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
* Fix plamo2 tokenizer session to prevent multiple calls of build()
---------
Co-authored-by: Francis Couture-Harpin <git@compilade.net>
Co-authored-by: Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>
* llama : deprecate llama_kv_self_ API
ggml-ci
* llama : allow llama_memory_(nullptr)
ggml-ci
* memory : add flag for optional data clear in llama_memory_clear
ggml-ci
* kv-cache : simplify the "struct llama_kv_cache" interface
ggml-ci
* kv-cache : revert the (n_swa + n_ubatch) change (for next PR)
ggml-ci
* kv-cache : some comments
ggml-ci
* context : fix graph reserve for multiple sequences
ggml-ci
* kv-cache : fix typo [no ci]
* kv-cache : fix find_slot() logic for free slots
ggml-ci
* llama : add TODO for deprecating the defrag API in the future
* kv-cache : improve find_slot() using min/max seq pos info
ggml-ci
* llama : handle aborts and compute errors
ggml-ci
* memory : extract state into llama_memory_state
ggml-ci
* kv-cache : add comments
ggml-ci
* server : update batching logic to reset n_batch on successful decode
* server : upon full re-processing, remove the sequence from the cache
* kv-cache : add TODO for doing split_equal when split_simple fails
ggml-ci
samuel
Johannes Gäßler
Aaron Teo
Sigbjørn Skjæret
Georgi Gerganov
Adrian Lundberg
JJJYmmm
Gadflyii
ddh0
Gabe Goodhart
Jonathan Graehl
Diego Devesa
Aman Gupta
Min-Hua
Shunta Saito
Xuan-Son Nguyen
Ed Addario
Ruikai Peng