[ROCm][Perf] Avoid extra reshape kernel in Qwen GDN output norm (#47842)

Signed-off-by: Mikko Tukiainen <Mikko.Tukiainen@amd.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
This commit is contained in:
Mikko Tukiainen
2026-09-22 14:38:16 +08:00
committed by GitHub
co-authored by Cursor
parent 496c6472cb
commit 68363d64ec
6 changed files with 170 additions and 33 deletions
@@ -852,13 +852,8 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention):
The RMSNormGated + quant sequence is eligible for fusion
by the compilation pass when fuse_norm_quant is enabled.
"""
z_shape_og = z.shape
core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
z = z.reshape(-1, z.shape[-1])
core_attn_out = self.norm(core_attn_out, z)
core_attn_out = core_attn_out.reshape(z_shape_og)
core_attn_out = core_attn_out.flatten(-2) # ... h d -> ... (h d)
output, _ = self.out_proj(core_attn_out)
output, _ = self.out_proj(core_attn_out.flatten(-2))
return output
def forward_hip(
@@ -1012,15 +1007,7 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention):
# ============================================================
# Part 3: Output Projection
# ============================================================
z_shape_og = z.shape
# Reshape input data into 2D tensor
core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
z = z.reshape(-1, z.shape[-1])
core_attn_out = self.norm(core_attn_out, z)
core_attn_out = core_attn_out.reshape(z_shape_og)
core_attn_out = core_attn_out.flatten(-2) # ... h d -> ... (h d)
out, _ = self.out_proj(core_attn_out)
return out
return self._output_projection(core_attn_out, z)
def forward_cpu(
self,
@@ -1063,14 +1050,7 @@ class QwenGatedDeltaNetAttention(GatedDeltaNetAttention):
_encode_layer_name(self.prefix),
)
z_shape_og = z.shape
core_attn_out = core_attn_out.reshape(-1, core_attn_out.shape[-1])
z = z.reshape(-1, z.shape[-1])
core_attn_out = self.norm(core_attn_out, z)
core_attn_out = core_attn_out.reshape(z_shape_og)
core_attn_out = core_attn_out.flatten(-2) # ... h d -> ... (h d)
out, _ = self.out_proj(core_attn_out)
return out
return self._output_projection(core_attn_out, z)
def _warmup_prefill_kernels(self, qkv_or_qkvz: torch.Tensor, v_dim: int) -> None:
"""Warm up GDN prefill kernels during V1 profiling.