fix: make OCR portable and reliable across AMD64 and ARM64 (#519)

* fix: make OCR portable and reliable

* fix: harden OCR installation portability

* fix: pin OCR partials across downloads

* fix: make OCR execution reliably asynchronous

* fix: harden OCR portability and docs routes

* fix: preserve decoder and docs safeguards

packages/ai/python/tests/test_gpu_detection.py

@@ -1,15 +1,5 @@
-"""gpu_available() must recognize a GPU that only paddle can use.
-
-The OCR feature bundle ships paddlepaddle-gpu but no torch or ONNX Runtime. On an
-OCR-only GPU host the torch and ONNX probes both come up empty, so before this
-fix gpu_available() fell through to a plain nvidia-smi check that returned False
-by design, and OCR silently downgraded to Tesseract. gpu_available() now probes
-paddle too, in an isolated subprocess, but only after nvidia-smi confirms a GPU
-is physically present (importing paddlepaddle-gpu in-process segfaults on a
-GPU-less host and would wedge the shared AI dispatcher).
-"""
+"""Framework-specific CUDA probes must never depend on obsolete Paddle OCR."""
 import os
-import subprocess
 import sys
 import types
 
@@ -17,32 +7,6 @@ sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
 import gpu  # noqa: E402
 
 
-# --- The isolated paddle probe --------------------------------------------
-
-def test_paddle_probe_true_when_subprocess_reports_cuda(monkeypatch):
-    def fake_run(cmd, **kwargs):
-        return subprocess.CompletedProcess(cmd, returncode=0, stdout="", stderr="")
-
-    monkeypatch.setattr(gpu.subprocess, "run", fake_run)
-    assert gpu._try_paddle_cuda_subprocess() is True
-
-
-def test_paddle_probe_false_when_subprocess_reports_no_cuda(monkeypatch):
-    def fake_run(cmd, **kwargs):
-        return subprocess.CompletedProcess(cmd, returncode=1, stdout="", stderr="")
-
-    monkeypatch.setattr(gpu.subprocess, "run", fake_run)
-    assert gpu._try_paddle_cuda_subprocess() is False
-
-
-def test_paddle_probe_false_on_timeout(monkeypatch):
-    def fake_run(cmd, **kwargs):
-        raise subprocess.TimeoutExpired(cmd, 30)
-
-    monkeypatch.setattr(gpu.subprocess, "run", fake_run)
-    assert gpu._try_paddle_cuda_subprocess() is False
-
-
 # --- gpu_available() orchestration -----------------------------------------
 
 def _patch_probes(monkeypatch, torch, onnx, smi):
@@ -54,39 +18,19 @@ def _patch_probes(monkeypatch, torch, onnx, smi):
     gpu.gpu_available.cache_clear()
 
 
-def test_gpu_available_true_when_only_paddle_sees_gpu(monkeypatch):
-    # OCR-only GPU box: torch and ONNX absent, GPU present, paddle can use it.
+def test_gpu_available_does_not_treat_hardware_presence_as_framework_support(monkeypatch):
     _patch_probes(monkeypatch, torch=False, onnx=False, smi="NVIDIA GeForce RTX 4070")
-    monkeypatch.setattr(gpu, "_try_paddle_cuda_subprocess", lambda: True)
-    assert gpu.gpu_available() is True
-
-
-def test_gpu_available_false_when_paddle_cannot_use_gpu(monkeypatch):
-    # GPU present but paddle is a CPU build or cannot init CUDA: stay conservative.
-    _patch_probes(monkeypatch, torch=False, onnx=False, smi="NVIDIA GeForce RTX 4070")
-    monkeypatch.setattr(gpu, "_try_paddle_cuda_subprocess", lambda: False)
     assert gpu.gpu_available() is False
 
 
-def test_gpu_available_never_probes_paddle_without_a_gpu(monkeypatch):
-    # Safety invariant: on a GPU-less host a paddlepaddle-gpu import segfaults, so
-    # the probe must never run when nvidia-smi finds no GPU.
-    _patch_probes(monkeypatch, torch=False, onnx=False, smi=None)
-    called = {"paddle": False}
-
-    def spy():
-        called["paddle"] = True
-        return True
-
-    monkeypatch.setattr(gpu, "_try_paddle_cuda_subprocess", spy)
-    assert gpu.gpu_available() is False
-    assert called["paddle"] is False
+def test_gpu_module_has_no_paddle_probe():
+    assert not hasattr(gpu, "_try_paddle_cuda_subprocess")
 
 
 # --- Per-framework detection (torch, ctranslate2) --------------------------
 #
 # Torch and CTranslate2 tools must gate on their OWN framework, not the general
-# gpu_available(), which can report True based on paddle or ONNX Runtime while
+# gpu_available(), which can report True based on ONNX Runtime while
 # torch is a CPU-only build. Consuming the shared boolean would make those tools
 # route to CUDA on a device their framework cannot use.
 
@@ -103,7 +47,7 @@ def test_torch_gpu_available_false_when_override_disables_gpu(monkeypatch):
 
 
 def test_torch_gpu_available_false_when_torch_is_cpu_only(monkeypatch):
-    # The crux: gpu_available() may be True via paddle or ONNX on a GPU box, but a
+    # The crux: gpu_available() may be True via ONNX on a GPU box, but a
     # CPU-only torch build must report no GPU so torch tools do not touch CUDA.
     monkeypatch.delenv("SNAPOTTER_GPU", raising=False)
     monkeypatch.setattr(gpu, "_try_torch_cuda", lambda: False)