Skip to content

LLM inference in C/C++

C++ ggml

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

Aman Gupta839a572a4cde
diff --git a/src/models/qwen35.cpp b/src/models/qwen35.cpp
index 802f2e2f13..8a4e434117 100644
--- a/src/models/qwen35.cpp
+++ b/src/models/qwen35.cpp
@@ -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;
}
diff --git a/src/models/qwen35moe.cpp b/src/models/qwen35moe.cpp
index fdbb2c63d4..939f10f7f0 100644
--- a/src/models/qwen35moe.cpp
+++ b/src/models/qwen35moe.cpp
@@ -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;
}
diff --git a/src/models/qwen3next.cpp b/src/models/qwen3next.cpp
index 13aa68ee80..98fafd917b 100644
--- a/src/models/qwen3next.cpp
+++ b/src/models/qwen3next.cpp
@@ -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;
}
diff --git a/src/models/qwen4exp.cpp b/src/models/qwen4exp.cpp
index 77430e7afd..22e7159840 100644
--- a/src/models/qwen4exp.cpp
+++ b/src/models/qwen4exp.cpp
@@ -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;
}