Skip to content

Load Sliding Window Attention (SWA) pattern from GGUF metadata (qwen3) #17597

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 4 commits into
base: master
Choose a base branch
from
+37 −5
Open

Load Sliding Window Attention (SWA) pattern from GGUF metadata (qwen3) #17597

Show file tree
Hide file tree
Changes from 2 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
29878b2
read swa pattern for all models
taylorchu Nov 29, 2025
5a0c228
add missing template
taylorchu Nov 29, 2025
b355891
bool is not supported
taylorchu Nov 29, 2025
cbcd6f9
qwen3 only
taylorchu Nov 30, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
1 change: 1 addition & 0 deletions src/llama-arch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -201,6 +201,7 @@ static const std::map<llm_kv, const char *> LLM_KV_NAMES = {
{ LLM_KV_ATTENTION_GATE_LORA_RANK, "%s.attention.gate_lora_rank" },
{ LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, "%s.attention.relative_buckets_count" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW, "%s.attention.sliding_window" },
{ LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, "%s.attention.sliding_window_pattern" },
{ LLM_KV_ATTENTION_SCALE, "%s.attention.scale" },
{ LLM_KV_ATTENTION_OUTPUT_SCALE, "%s.attention.output_scale" },
{ LLM_KV_ATTENTION_TEMPERATURE_LENGTH, "%s.attention.temperature_length" },
Expand Down
1 change: 1 addition & 0 deletions src/llama-arch.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -205,6 +205,7 @@ enum llm_kv {
LLM_KV_ATTENTION_GATE_LORA_RANK,
LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT,
LLM_KV_ATTENTION_SLIDING_WINDOW,
LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN,
LLM_KV_ATTENTION_SCALE,
LLM_KV_ATTENTION_OUTPUT_SCALE,
LLM_KV_ATTENTION_TEMPERATURE_LENGTH,
Expand Down
1 change: 1 addition & 0 deletions src/llama-model-loader.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -466,6 +466,7 @@ namespace GGUFMeta {
template bool llama_model_loader::get_key_or_arr<std::array<int, 4>>(enum llm_kv kid, std::array<int, 4> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<uint32_t, 512>>(enum llm_kv kid, std::array<uint32_t, 512> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<float, 512>>(enum llm_kv kid, std::array<float, 512> & result, uint32_t n, bool required);
template bool llama_model_loader::get_key_or_arr<std::array<bool, 512>>(enum llm_kv kid, std::array<bool, 512> & result, uint32_t n, bool required);


llama_model_loader::llama_model_loader(
Expand Down
1 change: 1 addition & 0 deletions src/llama-model-saver.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -182,6 +182,7 @@ void llama_model_saver::add_kv_from_model() {
add_kv(LLM_KV_ATTENTION_KV_LORA_RANK, hparams.n_lora_kv);
add_kv(LLM_KV_ATTENTION_RELATIVE_BUCKETS_COUNT, hparams.n_rel_attn_bkts);
add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa);
add_kv(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, true);
add_kv(LLM_KV_ATTENTION_SCALE, hparams.f_attention_scale);

const float rope_scaling_factor = hparams.rope_freq_scale_train == 1.0f ? 0.0f : 1.0f/hparams.rope_freq_scale_train;
Expand Down
8 changes: 8 additions & 0 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -555,6 +555,14 @@ void llama_model::load_hparams(llama_model_loader & ml) {

ml.get_key_or_arr(LLM_KV_ATTENTION_HEAD_COUNT_KV, hparams.n_head_kv_arr, hparams.n_layer, false);

ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
if (hparams.n_swa > 0) {
hparams.swa_type = LLAMA_SWA_TYPE_STANDARD;
Copy link
Collaborator

@ngxson ngxson Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't think we can assume anything about swa_type type here. Please remove this line

Copy link
Author

@taylorchu taylorchu Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This line is required; otherwise swa won't activate. This similar structure is used below.

Copy link
Collaborator

@ngxson ngxson Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I don't get it. how does swa not get activate?

swa_type should be set per-model. Please remove this line.

Copy link
Author

@taylorchu taylorchu Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more. 

        case LLAMA_SWA_TYPE_STANDARD:
            {
                if (p1 - p0 >= (int32_t) n_swa) {
                    return true;
                }
            } break;

swa requires specific masking to work. Without it, it will generate gibberish. The default semantic "standard". I think specific models can set to chunked or symmetric.

Copy link
Author

@taylorchu taylorchu Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

context: https://github.com/search?q=repo%3Ahuggingface%2Ftransformers+sliding_attention&type=code&p=4

ml.get_key_or_arr(LLM_KV_ATTENTION_SLIDING_WINDOW_PATTERN, hparams.swa_layers, hparams.n_layer, false);
} else {
hparams.swa_type = LLAMA_SWA_TYPE_NONE;
}
Copy link
Collaborator

@ngxson ngxson Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Remove this branch too, it's redundant

Copy link
Author

@taylorchu taylorchu Nov 29, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I can remove this because it is LLAMA_SWA_TYPE_NONE by default, but it seems like LLAMA_SWA_TYPE_NONE is added too below. I added it just in case.


bool rope_finetuned = false;
ml.get_key(LLM_KV_ROPE_SCALING_FINETUNED, rope_finetuned, false);
hparams.rope_finetuned = rope_finetuned;
Expand Down
Loading