Fix pre-commit lint issues (N806 + ruff format)

- Rename local `FP8_MAX` to `fp8_max` in `FP8QuantExporter._insert_conv_weight_dq_nodes`.
- Rename local `DQ_OPS` to `dq_ops` in `fold_dq_fp32_to_fp16_casts`.
- Apply ruff-format to collapse a broken-up chained expression in fp8_exporter.py.

Signed-off-by: ajrasane <131806219+ajrasane@users.noreply.github.com>

Author: ajrasane

modelopt/onnx/export/fp8_exporter.py

@@ -121,7 +121,7 @@ def _quantize_conv_weights_to_fp8(graph: gs.Graph) -> int:
         Returns:
             Number of Conv weight DQ nodes inserted.
         """
-        FP8_MAX = 448.0
+        fp8_max = 448.0
         count = 0
 
         for node in list(graph.nodes):
@@ -142,12 +142,10 @@ def _quantize_conv_weights_to_fp8(graph: gs.Graph) -> int:
             amax = torch_weights.abs().max().float()
             if amax == 0:
                 continue
-            scale_val = (amax / FP8_MAX).item()
+            scale_val = (amax / fp8_max).item()
 
             # Quantize weights to FP8 (WAR: numpy doesn't support fp8)
-            fp8_data = (
-                (torch_weights / scale_val).to(torch.float8_e4m3fn).view(torch.uint8).numpy()
-            )
+            fp8_data = (torch_weights / scale_val).to(torch.float8_e4m3fn).view(torch.uint8).numpy()
             fp8_tensor = onnx.TensorProto()
             fp8_tensor.data_type = onnx.TensorProto.FLOAT8E4M3FN
             fp8_tensor.dims.extend(fp8_data.shape)

modelopt/onnx/utils.py

@@ -1519,7 +1519,7 @@ def fold_dq_fp32_to_fp16_casts(onnx_model: onnx.ModelProto) -> onnx.ModelProto:
     Returns:
         The ONNX model with Cast nodes removed and DQ outputs set to FP16.
     """
-    DQ_OPS = {"DequantizeLinear", "TRT_FP8DequantizeLinear"}
+    dq_ops = {"DequantizeLinear", "TRT_FP8DequantizeLinear"}
 
     # Build a map of tensor name -> producer node
     producer_map: dict[str, onnx.NodeProto] = {}
@@ -1547,7 +1547,7 @@ def fold_dq_fp32_to_fp16_casts(onnx_model: onnx.ModelProto) -> onnx.ModelProto:
 
         # Check: producer is a DQ node
         producer = producer_map.get(node.input[0])
-        if producer is None or producer.op_type not in DQ_OPS:
+        if producer is None or producer.op_type not in dq_ops:
             continue
 
         # Convert the DQ scale initializer from FP32 to FP16