Torch TRT ngc container changes (#3299)

Co-authored-by: Dheeraj Peri <[email protected]>

core/util/Exception.cpp

@@ -1,3 +1,13 @@
+#if defined(__GNUC__) && !defined(__clang__)
+#if __GNUC__ >= 13
+#include <cstdint>
+#endif
+#elif defined(__clang__)
+#if __clang_major__ >= 13
+#include <cstdint>
+#endif
+#endif
+
 #include "core/util/Exception.h"
 
 #include <iostream>

noxfile.py

@@ -258,11 +258,12 @@ def run_dynamo_runtime_tests(session):
     tests = [
         "runtime",
     ]
+    skip_tests = "-k not hw_compat"
     for test in tests:
         if USE_HOST_DEPS:
-            session.run_always("pytest", test, env={"PYTHONPATH": PYT_PATH})
+            session.run_always("pytest", test, skip_tests, env={"PYTHONPATH": PYT_PATH})
         else:
-            session.run_always("pytest", test)
+            session.run_always("pytest", test, skip_tests)
 
 
 def run_dynamo_model_compile_tests(session):
@@ -332,7 +333,6 @@ def run_int8_accuracy_tests(session):
     tests = [
         "ptq/test_ptq_to_backend.py",
         "ptq/test_ptq_dataloader_calibrator.py",
-        "qat/",
     ]
     for test in tests:
         if USE_HOST_DEPS:
@@ -473,7 +473,6 @@ def run_l1_int8_accuracy_tests(session):
         install_deps(session)
         install_torch_trt(session)
     train_model(session)
-    finetune_model(session)
     run_int8_accuracy_tests(session)
     cleanup(session)
 

tests/py/dynamo/lowering/test_aten_lowering_passes.py

@@ -3,10 +3,17 @@
 
 import torch
 import torch_tensorrt
+from torch.testing._internal.common_cuda import PLATFORM_SUPPORTS_FLASH_ATTENTION
 from torch.testing._internal.common_utils import TestCase, run_tests
 
 from ..testing_utilities import DECIMALS_OF_AGREEMENT, lower_graph_testing
 
+isSM8XDevice = torch.cuda.is_available() and torch.cuda.get_device_capability() in [
+    (8, 6),
+    (8, 7),
+    (8, 9),
+]
+
 
 class TestInputAsOutput(TestCase):
     def test_input_as_output(self):
@@ -279,6 +286,10 @@ def forward(self, q, k, v):
     "Test not supported on Windows",
 )
 class TestLowerFlashAttention(TestCase):
+    @unittest.skipIf(
+        not PLATFORM_SUPPORTS_FLASH_ATTENTION or not isSM8XDevice,
+        "Does not support fused SDPA or not SM86+ hardware",
+    )
     def test_lower_flash_attention(self):
         class FlashAttention(torch.nn.Module):
             def forward(self, q, k, v):
@@ -348,6 +359,10 @@ def forward(self, q, k, v):
         )
         torch._dynamo.reset()
 
+    @unittest.skipIf(
+        not PLATFORM_SUPPORTS_FLASH_ATTENTION or not isSM8XDevice,
+        "Does not support fused SDPA or not SM86+ hardware",
+    )
     def test_flash_attention_converter(self):
         class FlashAttention(torch.nn.Module):
             def forward(self, q, k, v):