qwen: reshape output to 2d instead of 3d (#30823)

* qwen: reshape output to 2d instead of 3d

* remove final reshape, fix comments
This commit is contained in:
Aman Gupta authored and GitHub committed 2026-10-20 22:09:56 +03:00
1 parent ead9f32058
commit 839a572a4c
4 files changed
+7 -19

No files matched your search

+2 -5
View File
@@ -465,17 +465,14 @@ ggml_tensor * llama_model_qwen35::graph::build_layer_attn_linear(
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * head_dim , n_tokens * n_seqs ]
ggml_tensor * final_output = ggml_reshape_2d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens * n_seqs);
cb(final_output, "final_output", il);
// Output projection
cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s);
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
+2 -6
View File
@@ -475,17 +475,13 @@ ggml_tensor * llama_model_qwen35moe::graph::build_layer_attn_linear(
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
cb(final_output, "final_output", il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * head_dim , n_tokens * n_seqs ]
ggml_tensor * final_output = ggml_reshape_2d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens * n_seqs);
// Output projection
cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s);
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
+2 -5
View File
@@ -542,17 +542,14 @@ ggml_tensor * llama_model_qwen3next::graph::build_layer_attn_linear(
// Apply gated normalization: self.norm(core_attn_out, z)
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_tokens, n_seqs, n_heads * head_dim]
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
// Final reshape: [head_dim, n_heads, n_tokens, n_seqs] -> [n_heads * head_dim , n_tokens * n_seqs ]
ggml_tensor * final_output = ggml_reshape_2d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens * n_seqs);
cb(final_output, "final_output", il);
// Output projection
cur = build_lora_mm(model.layers[il].ssm_out, final_output);
cb(cur, "linear_attn_out", il);
// Reshape back to original dimensions
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}
+1 -3
View File
@@ -1131,14 +1131,12 @@ ggml_tensor * llama_model_qwen4exp::graph::build_layer_attn_linear(
// gated normalization, as self.norm(core_attn_out, z) in the reference
ggml_tensor * attn_out_norm = build_norm_gated(output, model.layers[il].ssm_norm, z_2d, il);
ggml_tensor * final_output = ggml_reshape_3d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens, n_seqs);
ggml_tensor * final_output = ggml_reshape_2d(ctx0, attn_out_norm, head_v_dim * num_v_heads, n_seq_tokens * n_seqs);
cb(final_output, "final_output", il);
cur = build_lora_mm(model.layers[il].ssm_out, final_output, model.layers[il].ssm_out_s);
cb(cur, "linear_attn_out", il);
cur = ggml_reshape_2d(ctx0, cur, n_embd, n_seq_tokens * n_seqs);
return cur;
}