Skip to content

Fix kq_scale for the attention layers of PLaMo2 #14892

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
merged 7 commits into from
Jul 27, 2025
Merged

Fix kq_scale for the attention layers of PLaMo2 #14892

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
7 commits
Select commit Hold shift + click to select a range
7baf4fd
Fix dimensions for expand
mitmul Jul 18, 2025
e39bc09
Change dimensions to copy states to cache
mitmul Jul 26, 2025
bd4d2e1
Fix the default value for plamo2 conversion
mitmul Jul 26, 2025
c475203
Fix scale given to build_attn
mitmul Jul 26, 2025
75f0a0d
Update src/llama-model.cpp
mitmul Jul 27, 2025
429639d
Update src/llama-model.cpp
mitmul Jul 27, 2025
60a705d
Update src/llama-model.cpp
mitmul 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 @@ -3788,7 +3788,7 @@ 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 @@ -3799,7 +3799,7 @@ 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
15 changes: 8 additions & 7 deletions src/llama-model.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -15988,7 +15988,7 @@ 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 @@ -16028,7 +16028,7 @@ 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 @@ -16103,8 +16103,9 @@ 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 @@ -16167,9 +16168,9 @@ 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))));
Copy link
Collaborator

Choose a reason for hiding this comment

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

Mind commenting on these changes?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

These changes actually replace d_inner with n_heads*head_dim for both of y_ssm and ssm_states_all. This is because I thought it's more natural because the ssm state created in get_ssm_rows() has the shape of (d_state , head_dim, n_heads, n_seqs) here:

llama.cpp/src/llama-model.cpp

Line 16359 in 446595b

ggml_tensor * ssm = ggml_reshape_4d(ctx, states, d_state, head_dim, n_heads, mctx_cur->get_size());

But, if d_inner is always the same as n_heads*head_dim, I'm happy to revert this change cause it's unnecessary.

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