From c22d5e8e8463af96ddce2a999c56bdabbe40e591 Mon Sep 17 00:00:00 2001
From: Chao Wang <26245345+ChaoWao@users.noreply.github.com>
Date: Sat, 11 Jul 2026 10:46:13 +0000
Subject: [PATCH] Fix: compile host_build_graph orchestration with host g++
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

host_build_graph dlopens the orchestration .so on the x86_64 host, but
compile_orchestration routed a2a3/a5 to the aarch64 cross-compiler
unconditionally — that routing predates host_build_graph. The AArch64
orch .so could not be dlopened by the x86_64 host runtime, surfacing as
"dlopen failed: ... No such file or directory" and failing every
host_build_graph scene test on a2a3 onboard.

Route the orch toolchain by runtime instead of only by platform:
host_build_graph uses HOST_GXX (the host-dlopen path, onboard and sim);
tensormap_and_ringbuffer keeps AARCH64_GXX onboard (AICPU dlopen) and
HOST_GXX for sim. Unknown runtime_name now raises ValueError, matching
the docstring's documented contract.

Verified on a2a3 onboard (910B2C): host_build_graph vector_example and
paged_attention now pass; tensormap alternating_matmul_add unchanged
(regression guard). a5 host_build_graph takes the same path but has no
a5 silicon here to onboard-verify.
---
 simpler_setup/kernel_compiler.py | 30 ++++++++++++++++++------------
 1 file changed, 18 insertions(+), 12 deletions(-)

diff --git a/simpler_setup/kernel_compiler.py b/simpler_setup/kernel_compiler.py
index b2aff3b5a0..96cdf72598 100644
--- a/simpler_setup/kernel_compiler.py
+++ b/simpler_setup/kernel_compiler.py
@@ -419,15 +419,23 @@ def compile_orchestration(
         if orch_includes:
             include_dirs = include_dirs + orch_includes
 
-        # Resolve toolchain: HOST_GXX needs no runtime-specific extras
-        toolchain_type = self._get_toolchain(
-            {
-                "a2a3": ToolchainType.AARCH64_GXX,
-                "a2a3sim": ToolchainType.HOST_GXX,
-                "a5": ToolchainType.AARCH64_GXX,
-                "a5sim": ToolchainType.HOST_GXX,
-            },
-        )
+        # host_build_graph dlopens the orchestration .so on the host, so it
+        # compiles with the host g++ (x86_64) regardless of platform.
+        # tensormap_and_ringbuffer dlopens it on the aarch64 AICPU onboard, so
+        # it cross-compiles there and uses the host g++ only for sim.
+        if runtime_name == "host_build_graph":
+            toolchain_type = ToolchainType.HOST_GXX
+        elif runtime_name == "tensormap_and_ringbuffer":
+            toolchain_type = self._get_toolchain(
+                {
+                    "a2a3": ToolchainType.AARCH64_GXX,
+                    "a2a3sim": ToolchainType.HOST_GXX,
+                    "a5": ToolchainType.AARCH64_GXX,
+                    "a5sim": ToolchainType.HOST_GXX,
+                },
+            )
+        else:
+            raise ValueError(f"Unknown runtime_name: {runtime_name!r}")
         toolchain: Union[GxxToolchain, Aarch64GxxToolchain]
         if toolchain_type == ToolchainType.AARCH64_GXX:
             assert self.aarch64 is not None, "aarch64 toolchain is only available for hardware platforms"
@@ -435,9 +443,7 @@ def compile_orchestration(
         else:
             toolchain = self.host_gxx
 
-        # HOST_GXX: simulation build (host execution)
-        # AARCH64_GXX: cross-compilation for supported runtimes
-        #   Note: orchestration uses ops table via pto_orchestration_api.h (no extra runtime sources needed)
+        # Orchestration uses the ops table via pto_orchestration_api.h (no extra runtime sources needed).
         return self._compile_orchestration_shared_lib(
             source_path,
             toolchain,