fix(hwfit): normalize CPU arch for fallback estimates (#4441)

tests/test_hwfit_cpu_arch_detection.py

@@ -0,0 +1,55 @@
+"""CPU architecture normalization for HW Fit hardware detection."""
+
+import pytest
+
+from services.hwfit import hardware
+
+
+@pytest.fixture(autouse=True)
+def _clear_hwfit_cache(monkeypatch):
+    hardware._cache_by_host.clear()
+    monkeypatch.setattr(hardware, "_remote_host", None)
+    monkeypatch.setattr(hardware, "_remote_platform", None)
+    monkeypatch.setattr(hardware, "_is_containerized", lambda: False)
+    yield
+    hardware._cache_by_host.clear()
+
+
+def _stub_common_probe(monkeypatch, machine):
+    monkeypatch.setattr(hardware.platform, "machine", lambda: machine)
+    monkeypatch.setattr(hardware, "_get_ram_gb", lambda: 64.0)
+    monkeypatch.setattr(hardware, "_get_available_ram_gb", lambda: 48.0)
+    monkeypatch.setattr(hardware, "_get_cpu_count", lambda: 16)
+    monkeypatch.setattr(hardware, "_get_cpu_name", lambda: "Test CPU")
+    monkeypatch.setattr(hardware, "_detect_apple_silicon", lambda: None)
+    monkeypatch.setattr(hardware, "_detect_amd", lambda: None)
+
+
+def test_detect_system_reports_cpu_arch_for_gpu_backends(monkeypatch):
+    """GPU-backed systems still need CPU architecture for cpu_only estimates."""
+    _stub_common_probe(monkeypatch, "aarch64")
+    monkeypatch.setattr(hardware, "_detect_nvidia", lambda: {
+        "gpu_name": "NVIDIA GB10",
+        "gpu_vram_gb": 64.0,
+        "gpu_count": 1,
+        "gpus": [],
+        "gpu_groups": [],
+        "homogeneous": True,
+        "backend": "cuda",
+    })
+
+    system = hardware.detect_system(fresh=True)
+
+    assert system["backend"] == "cuda"
+    assert system["cpu_arch"] == "arm64"
+
+
+def test_detect_system_keeps_32_bit_arm_on_conservative_cpu_backend(monkeypatch):
+    """Plain arm/armv7 is not the same as the ARM64-class cpu_arm fallback."""
+    _stub_common_probe(monkeypatch, "armv7l")
+    monkeypatch.setattr(hardware, "_detect_nvidia", lambda: None)
+
+    system = hardware.detect_system(fresh=True)
+
+    assert system["cpu_arch"] == "arm"
+    assert system["backend"] == "cpu_x86"

tests/test_hwfit_cpu_only_fallback.py

@@ -47,6 +47,12 @@ ARM64_SYSTEM = {
     "gpu_vram_gb": 0,
 }
 
+ARM32_SYSTEM = {
+    "backend": "arm",
+    "gpu_name": None,
+    "gpu_vram_gb": 0,
+}
+
 AARCH64_SYSTEM = {
     "backend": "aarch64",
     "gpu_name": None,
@@ -79,6 +85,16 @@ def test_cpu_only_on_metal_apple_silicon_uses_cpu_arm_fallback():
     assert metal_tps > 0
 
 
+def test_cpu_only_on_gpu_backend_uses_detected_arm64_cpu_arch():
+    """A GPU backend on an ARM64 host should use the ARM CPU fallback for cpu_only."""
+    cuda_arm64 = dict(CUDA_SYSTEM, cpu_arch="aarch64", cpu_name="Ampere Altra")
+    cuda_arm64_tps = _estimate_speed(DENSE_MODEL, QUANT, "cpu_only", cuda_arm64)
+    arm_tps = _estimate_speed(DENSE_MODEL, QUANT, "cpu_only", CPU_ARM_SYSTEM)
+
+    assert cuda_arm64_tps == pytest.approx(arm_tps, rel=1e-9, abs=1e-9)
+    assert cuda_arm64_tps > 0
+
+
 @pytest.mark.parametrize(
     "arm_alias_system",
     [ARM64_SYSTEM, AARCH64_SYSTEM, CPU_ARM_SYSTEM],
@@ -93,6 +109,15 @@ def test_cpu_only_preserves_arm_backends(arm_alias_system):
     assert alias_tps > 0
 
 
+def test_cpu_only_does_not_treat_plain_arm_as_arm64_fallback():
+    """Docker/OCI plain arm is not the ARM64-class fallback used for Apple Silicon."""
+    arm32_tps = _estimate_speed(DENSE_MODEL, QUANT, "cpu_only", ARM32_SYSTEM)
+    x86_tps = _estimate_speed(DENSE_MODEL, QUANT, "cpu_only", CPU_X86_SYSTEM)
+
+    assert arm32_tps == pytest.approx(x86_tps, rel=1e-9, abs=1e-9)
+    assert arm32_tps > 0
+
+
 def test_cpu_only_preserves_known_cpu_backends():
     """Known CPU backends should be preserved, not rewritten to cpu_x86."""
     for system in (CPU_X86_SYSTEM, CPU_ARM_SYSTEM):

tests/test_hwfit_macos.py

@@ -165,6 +165,15 @@ def test_intel_mac_skipped(monkeypatch):
     assert hardware._detect_apple_silicon() is None
 
 
+def test_plain_arm_mac_skipped(monkeypatch):
+    """Only ARM64-class Macs should enter the Apple Silicon Metal path."""
+    monkeypatch.setattr(hardware, "_remote_host", None)
+    monkeypatch.setattr(hardware.platform, "system", lambda: "Darwin")
+    monkeypatch.setattr(hardware.platform, "machine", lambda: "armv7l")
+    monkeypatch.setattr(hardware, "_run", _fake_sysctl())
+    assert hardware._detect_apple_silicon() is None
+
+
 def test_detect_system_propagates_unified_memory(monkeypatch):
     """The unified_memory flag set by GPU detection must survive into the
     system dict so the API and UI can report it (it was being dropped)."""