Skip redundant decode FFN operators when swiglu_fused is active

When use_aie_ffn_swiglu_fused is True, the fused operator handles all
of gate+up+silu+mul+down in one design. Skip creating the 3 separate
decode GEMVs (aie_fc1_gemv, aie_fc2_gemv, aie_fc3_gemv) which would
waste compilation time and device memory.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: jgmelber

iron/applications/llama_3.2_1b/src/block/feed_forward.py

@@ -123,7 +123,12 @@ def __init__(
                 cfg["hidden_dim"], cfg["emb_dim"], dtype=cfg["dtype"], bias=False
             )
 
-        if self.cfg["use_kv_cache"] and self.cfg["use_aie_ffn_gemv"]:
+        # Skip creating separate decode GEMVs when fused SwiGLU handles everything
+        if (
+            self.cfg["use_kv_cache"]
+            and self.cfg["use_aie_ffn_gemv"]
+            and not self.cfg.get("use_aie_ffn_swiglu_fused", False)
+        ):
             aie_gemv_config = {"num_aie_columns": 8, "is_mv": False}
             # FC1 and FC2: emb_dim -> hidden_dim
             self.aie_fc1_gemv = AIEGEMV(

iron/applications/llama_3.2_1b/src/model_with_json.py

@@ -94,7 +94,9 @@ def format_option(name, value):
             "use_aie_ffn_mul",
             "use_aie_ffn_silu",
         }
-        if not cfg.get("use_aie_ffn_swiglu_fused", False):
+        if cfg.get("use_aie_ffn_swiglu_fused", False):
+            dont_print |= {"use_aie_ffn_gemv"}
+        else:
             dont_print |= {"use_aie_ffn_swiglu_fused"}
     else:
         dont_print |= {"use_aie_ffn_swiglu", "use_aie_ffn_swiglu_fused"}