Skip to content

models: fix linear attention state corruption in recurrent layers #22661

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
aagit wants to merge 1 commit into
base: master
Choose a base branch
from
+15 −8
Open

models: fix linear attention state corruption in recurrent layers #22661

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
5 changes: 3 additions & 2 deletions src/models/kimi-linear.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -115,8 +115,6 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = build_norm(inpL, layer.attn_norm, NULL, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);

ggml_build_forward_expand(gf, cur);

if (hparams.is_recurrent(il)) {
// === KDA Layer (Kimi Delta Attention) with Recurrent State ===
// Reference: vLLM kda.py
Expand Down Expand Up @@ -203,7 +201,10 @@ llm_build_kimi_linear::llm_build_kimi_linear(const llama_model & model, const ll
cur = ggml_mul_mat(ctx0, layer.wo, gated);
cb(cur, "kda_out", il);

ggml_build_forward_expand(gf, cur);
} else {
ggml_build_forward_expand(gf, cur);

// === MLA Layer (Multi-head Latent Attention) without KV Cache ===
// Reference: vLLM mla.py
// Step 1: Q projection and reshape
Expand Down
6 changes: 4 additions & 2 deletions src/models/qwen35.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,13 +29,15 @@ llm_build_qwen35::llm_build_qwen35(const llama_model & model, const llm_graph_pa
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);

ggml_build_forward_expand(gf, cur);

// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);

ggml_build_forward_expand(gf, cur);
} else {
ggml_build_forward_expand(gf, cur);

// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
}
Expand Down
6 changes: 4 additions & 2 deletions src/models/qwen35moe.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,13 +29,15 @@ llm_build_qwen35moe::llm_build_qwen35moe(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);

ggml_build_forward_expand(gf, cur);

// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);

ggml_build_forward_expand(gf, cur);
} else {
ggml_build_forward_expand(gf, cur);

// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, sections, il);
}
Expand Down
6 changes: 4 additions & 2 deletions src/models/qwen3next.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -21,13 +21,15 @@ llm_build_qwen3next::llm_build_qwen3next(const llama_model & model, const llm_gr
cur = build_norm(inpL, model.layers[il].attn_norm, nullptr, LLM_NORM_RMS, il);
cb(cur, "attn_norm", il);

ggml_build_forward_expand(gf, cur);

// Determine layer type and build appropriate attention mechanism
if (hparams.is_recurrent(il)) {
// Linear attention layer (gated delta net)
cur = build_layer_attn_linear(inp->get_recr(), cur, il);

ggml_build_forward_expand(gf, cur);
} else {
ggml_build_forward_expand(gf, cur);

// Full attention layer
cur = build_layer_attn(inp->get_attn(), cur, inp_pos, il);
}
Expand Down