Changes to TRT-LLM download tool for multigpu distributed case

[apbose]

TRT-LLM installation tool for distributed

1. The download is to be done by only one GPU to avoid unnecessary downloads
2. Use of lock files in the tool for the above purpose

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/utils.py	2025-09-22 06:35:28.523784+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/utils.py	2025-09-22 06:36:00.657186+00:00
@@ -863,6 +863,6 @@
    return False


def is_thor() -> bool:
    if torch.cuda.get_device_capability() in [(11, 0)]:
-        return True
\ No newline at end of file
+        return True

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/distributed/utils.py	2025-09-25 19:33:28.176615+00:00
+++ /home/runner/work/TensorRT/TensorRT/py/torch_tensorrt/dynamo/distributed/utils.py	2025-09-25 19:34:02.325958+00:00
@@ -100,11 +100,10 @@
        return True

    except Exception as e:
        logger.warning(f"Failed to detect CUDA version: {e}")
        return False
-

    return True


def _cache_root() -> Path:

[github-actions]

There are some changes that do not conform to Python style guidelines:

--- /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/distributed/test_nccl_ops.py	2025-10-13 18:42:31.890493+00:00
+++ /home/runner/work/TensorRT/TensorRT/tests/py/dynamo/distributed/test_nccl_ops.py	2025-10-13 18:43:14.026641+00:00
@@ -23,10 +23,11 @@
if not dist.is_initialized():
    dist.init_process_group(
        backend="nccl",
        init_method="env://",
    )
+

class DistributedGatherModel(nn.Module):
    def __init__(self, input_dim, world_size, group_name):
        super().__init__()
        self.fc = nn.Linear(input_dim, input_dim)

[narendasan]

Why would it matter what the examples need? This is supposed to be user facing code.

[narendasan]

Where is this logger used? Typically this is a user's responsibility. we should not be creating the actual logger handlers in our library. We can do this in the examples code tho

[apbose]

ok was meant for the use case that if we want separate log outputs for the two different GPUs. But I agree these all can be moved to the user facing code