[AMD][CI] Pass USE_PDL explicitly in the fused MoE gate (#36306)

This commit is contained in:
Bingxu Chen
2026-08-25 19:57:59 -07:00
committed by GitHub
parent fa3ac61661
commit 5bcc978f2a
@@ -534,7 +534,8 @@ def fused_gate_sigmoid_mul_add(
if num_tokens >= 1024: if num_tokens >= 1024:
config["num_warps"] = min(config["num_warps"], 8) config["num_warps"] = min(config["num_warps"], 8)
pdl_kwargs = {"USE_PDL": True, "launch_pdl": True} if is_arch_support_pdl() else {} use_pdl = is_arch_support_pdl()
pdl_kwargs = {"launch_pdl": True} if use_pdl else {}
_fused_gate_sigmoid_mul_add_kernel[(num_tokens,)]( _fused_gate_sigmoid_mul_add_kernel[(num_tokens,)](
hidden_states, hidden_states,
@@ -542,6 +543,7 @@ def fused_gate_sigmoid_mul_add(
shared_output, shared_output,
final_hidden_states, final_hidden_states,
hidden_dim=hidden_dim, hidden_dim=hidden_dim,
USE_PDL=use_pdl,
**config, **config,
**pdl_kwargs, **pdl_kwargs,
) )