Skip to content

[pull] master from ggml-org:master #206

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

Merged
pull merged 2 commits into from
Jul 27, 2025
Merged

[pull] master from ggml-org:master #206

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
1dc9614
llama : fix kq_scale for the attention layers of PLaMo2 (#14892)
mitmul Jul 27, 2025
4762ad7
model : make rope_yarn_log_mul optional for deepseek2 (#14896)
gabriellarson Jul 27, 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
4 changes: 2 additions & 2 deletions convert_hf_to_gguf.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3791,7 +3791,7 @@ def set_gguf_parameters(self):
self.gguf_writer.add_block_count(block_count)
self.gguf_writer.add_head_count(hparams.get("num_attention_heads", 32))
self.gguf_writer.add_layer_norm_rms_eps(hparams.get("rms_norm_eps", 1e-06))
self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 1000000.0))
self.gguf_writer.add_rope_freq_base(hparams.get("rope_theta", 10000))

# Mamba parameters
self.gguf_writer.add_ssm_state_size(hparams.get("mamba_d_state", 64))
Expand All @@ -3802,7 +3802,7 @@ def set_gguf_parameters(self):
self.gguf_writer.add_ssm_group_count(0)

# MLP feed forward parameters (for attention layers)
self.gguf_writer.add_feed_forward_length(hparams.get("intermediate_size", 16384))
self.gguf_writer.add_feed_forward_length(hparams.get("intermediate_size", 13312))
self.gguf_writer.add_file_type(self.ftype)

def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
Expand Down
2 changes: 1 addition & 1 deletion src/llama-hparams.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -98,7 +98,7 @@ struct llama_hparams {
float rope_freq_scale_train;
float rope_freq_scale_train_swa;
uint32_t n_ctx_orig_yarn;
float rope_yarn_log_mul;
float rope_yarn_log_mul = 0.0f;

std::array<int, 4> rope_sections;

Expand Down
17 changes: 9 additions & 8 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1369,7 +1369,7 @@ void llama_model::load_hparams(llama_model_loader & ml) {
// that have no expert_gating_func model parameter set
hparams.expert_gating_func = LLAMA_EXPERT_GATING_FUNC_TYPE_SOFTMAX;
}
ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul);
ml.get_key(LLM_KV_ROPE_SCALING_YARN_LOG_MUL, hparams.rope_yarn_log_mul, false);

switch (hparams.n_layer) {
case 27: type = LLM_TYPE_16B; break;
Expand Down Expand Up @@ -16191,7 +16191,7 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
{
// PLaMo-2 uses combined QKV tensor
ggml_tensor * qkv = build_lora_mm(model.layers[il].wqkv, cur);
cb(qkv, "qkv", il);
cb(qkv, "wqkv", il);

// split QKV tensor into Q, K, V
const int64_t n_embd_head_q = hparams.n_embd_head_k;
Expand Down Expand Up @@ -16231,7 +16231,7 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
ext_factor, attn_factor, beta_fast, beta_slow
);

cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f, il);
cur = build_attn(inp, model.layers[il].wo, NULL, Qcur, Kcur, Vcur, NULL, NULL, 1.0f/sqrtf(float(n_embd_head_v)), il);
}

cb(cur, "attn_out", il);
Expand Down Expand Up @@ -16306,8 +16306,9 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
ggml_build_forward_expand(gf,
ggml_cpy(ctx0, last_conv,
ggml_view_1d(ctx0, conv_states_all,
(d_conv - 1)*(d_inner)*(n_seqs),
kv_head*(d_conv - 1)*(d_inner)*ggml_element_size(conv_states_all))));
(d_conv - 1)*(d_inner + 2*n_group*d_state)*(n_seqs),
kv_head*(d_conv - 1)*(d_inner + 2*n_group*d_state)*ggml_element_size(conv_states_all))));
cb(conv_states_all, "mamba_conv1d_state", il);

// 1D convolution
x = ggml_ssm_conv(ctx0, conv_x, model.layers[il].ssm_conv1d);
Expand Down Expand Up @@ -16370,9 +16371,9 @@ struct llm_build_plamo2 : public llm_graph_context_mamba {
// store last states
ggml_build_forward_expand(gf,
ggml_cpy(ctx0,
ggml_view_1d(ctx0, y_ssm, d_state*d_inner*n_seqs, x->nb[3]*x->ne[3]),
ggml_view_1d(ctx0, ssm_states_all, d_state*d_inner*n_seqs,
kv_head*d_state*d_inner*ggml_element_size(ssm_states_all))));
ggml_view_1d(ctx0, y_ssm, n_heads*head_dim*d_state*n_seqs, n_heads*head_dim*n_seq_tokens*n_seqs*ggml_element_size(y_ssm)),
ggml_view_1d(ctx0, ssm_states_all, n_heads*head_dim*d_state*n_seqs, kv_head*n_seqs*n_heads*head_dim*d_state*ggml_element_size(ssm_states_all))));
cb(ssm_states_all, "mamba_ssm_states", il);

ggml_tensor * y = ggml_view_4d(ctx0, y_ssm, head_dim, n_heads, n_seq_tokens, n_seqs, head_dim * ggml_element_size(x), head_dim * n_heads * ggml_element_size(x), head_dim * n_heads * n_seq_tokens * ggml_element_size(x), 0);
cb(y, "mamba_y_view", il);
Expand Down
Loading