Added column-major storage of weights and scales in INT4 quantization for model load time improvement in TRT-RTX

examples/windows/onnx_ptq/genai_llm/README.md

@@ -60,6 +60,7 @@ The table below lists key command-line arguments of the ONNX PTQ example script.
 | `--layers_8bit` | Default: None | Use this option to Overrides default mixed quant strategy|
 | `--gather_quantize_axis` | Default: None | Use this option to enable INT4 quantization of Gather nodes - choose 0 or 1|
 | `--gather_block_size` | Default: 32 | Block-size for Gather node's INT4 quantization (when its enabled using gather_quantize_axis option)|
+| `--use_column_major` | Default: disabled | Apply column-major storage optimization for execution providers that need it. Only applicable for DQ-only quantization.|
 
 Run the following command to view all available parameters in the script:
 

examples/windows/onnx_ptq/genai_llm/quantize.py

@@ -369,7 +369,7 @@ def main(args):
         f"batch_size={args.batch_size}, block_size={args.block_size}, add-position-ids={args.add_position_ids}, "
         f"past-kv={args.add_past_kv_inputs}, rcalib={args.use_random_calib}, device={args.device}, "
         f"use_zero_point={args.use_zero_point}, use_fp32={args.use_fp32} enable_mixed_quant={args.enable_mixed_quant}, "
-        f"layers_8bit={args.layers_8bit}\n"
+        f"layers_8bit={args.layers_8bit}, use_column_major={args.use_column_major}\n"
     )
 
     print(
@@ -443,6 +443,7 @@ def main(args):
         layers_8bit=args.layers_8bit,
         gather_block_size=args.gather_block_size,
         gather_quantize_axis=args.gather_quantize_axis,
+        use_column_major=args.use_column_major,
     )
     logging.info(f"\nQuantization process took {time.time() - t} seconds")
 
@@ -629,5 +630,14 @@ def main(args):
         default="",
         help=("Overrides default mixed quant strategy. Example: 'layers.0,lm_head'"),
     )
+    parser.add_argument(
+        "--use_column_major",
+        default=False,
+        action="store_true",
+        help=(
+            "Apply column-major storage optimization for execution providers that need it. "
+            "Only applicable for DQ-only quantization (e.g., rtn_dq, awq_lite, awq_clip)."
+        ),
+    )
     args = parser.parse_args()
     main(args)

modelopt/onnx/quantization/int4.py

@@ -220,7 +220,21 @@ def quantize_rtn(
 
     Always selects the first dimension (0) to block over. This is because we must batch over the Cin
     dimension, and in ONNX, weights are always plugged into the RHS (i.e. y = x @ W).
+
+    Args:
+        use_column_major: If True, apply column-major storage optimization for execution
+                          providers that need it. Passed via kwargs.
     """
+    use_column_major = kwargs.get("use_column_major", False)
+
+    # Column-major only makes sense for DQ-only mode
+    if use_column_major and not dq_only:
+        logger.warning(
+            "use_column_major=True has no effect in QDQ mode. "
+            "Column-major optimization only applies to DQ-only quantization."
+        )
+        use_column_major = False
+
     logger.info("Starting RTN quantization")
     t_start = time.time()
 
@@ -295,8 +309,15 @@ def quantize_rtn(
                 qw = np.asnumpy(qw)
                 scales[name] = np.asnumpy(scales[name])
             gemm_weights_quantized[name] = numpy.asarray(qw)
+        # Apply column-major optimization if flag is set
+        # Transposes the weights and scales in-place
+        if use_column_major:
+            qdq.apply_column_major_transformation(gemm_weights_quantized, scales)
+            dq_node_attributes = {"axis": 1, "block_size": block_size}
+        else:
+            dq_node_attributes = {"axis": 0, "block_size": block_size}
+
         scales = reshape_scales_for_per_channel_nodes(scales, block_size, layer_info)
-        dq_node_attributes = {"axis": 0, "block_size": block_size}
         qdq.insert_dq_nodes(
             graph,
             scales,
@@ -305,6 +326,10 @@ def quantize_rtn(
             layer_info=layer_info,
         )
 
+        # Add transpose nodes for column-major if needed
+        if use_column_major:
+            qdq.insert_transpose_nodes_for_column_major(graph)
+
         if gather_w_map is not None:
             gather_dq_node_attributes = {
                 "axis": gather_quantize_axis,
@@ -605,7 +630,14 @@ def _quantize_awq_clip(
         )
 
     t = time.time()
-    dq_node_attributes = {"axis": 0, "block_size": block_size}
+    # Apply column-major optimization if flag is set
+    # Transposes the weights and scales in-place
+    use_column_major = kwargs.get("use_column_major", False)
+    if use_column_major:
+        qdq.apply_column_major_transformation(gemm_weights_quantized, scales)
+        dq_node_attributes = {"axis": 1, "block_size": block_size}
+    else:
+        dq_node_attributes = {"axis": 0, "block_size": block_size}
     scales = reshape_scales_for_per_channel_nodes(scales, block_size, layer_info)
     qdq.insert_dq_nodes(
         graph_gs,
@@ -614,6 +646,9 @@ def _quantize_awq_clip(
         attributes=dq_node_attributes,
         layer_info=layer_info,
     )
+    # Add transpose nodes for column-major if needed
+    if use_column_major:
+        qdq.insert_transpose_nodes_for_column_major(graph_gs)
     if gather_w_map is not None:
         assert gather_s_map is not None, "scale-map not found for quantizable gather nodes"
         gather_dq_node_attributes = {"axis": gather_quantize_axis, "block_size": gather_block_size}
@@ -1308,7 +1343,14 @@ def _quantize_awq_lite(
         )
 
     t = time.time()
-    dq_node_attributes = {"axis": 0, "block_size": block_size}
+    # Apply column-major optimization if flag is set
+    # Transposes the weights and scales in-place
+    use_column_major = kwargs.get("use_column_major", False)
+    if use_column_major:
+        qdq.apply_column_major_transformation(gemm_weights_quantized, scales)
+        dq_node_attributes = {"axis": 1, "block_size": block_size}
+    else:
+        dq_node_attributes = {"axis": 0, "block_size": block_size}
     scales = reshape_scales_for_per_channel_nodes(scales, block_size, layer_info)
     qdq.insert_dq_nodes(
         graph_gs,
@@ -1318,6 +1360,9 @@ def _quantize_awq_lite(
         zero_points=zero_points if use_zero_point else None,
         layer_info=layer_info,
     )
+    # Add transpose nodes for column-major if needed
+    if use_column_major:
+        qdq.insert_transpose_nodes_for_column_major(graph_gs)
     if gather_w_map is not None:
         assert gather_s_map is not None, "scale-map not found for quantizable gather nodes"
         assert not use_zero_point or gather_zp_map, (
@@ -1420,10 +1465,19 @@ def quantize(
                                               Default: False.
                 - **layers_8bit** (str): comma-separated list of layer patterns to quantize to INT8 instead of INT4.
                                               Default: [].
+                - **use_column_major** (bool): If True, apply column-major storage optimization for
+                                              execution providers that need it. This transposes
+                                              weights and adds Transpose nodes around MatMul operations.
+                                              Only applies to DQ-only quantization mode.
+                                              Default: False.
     **Returns**: A quantized ONNX model in ONNX ModelProto format.
     """
     configure_logging(level=log_level.upper())
     logger.info(f"Starting INT4 quantization with method: {calibration_method}")
+
+    # Log if column-major optimization is enabled (works for all methods)
+    if kwargs.get("use_column_major", False):
+        logger.info("Column-major storage optimization enabled via use_column_major flag")
     t_start = time.time()
 
     if cupy_warning_msg:

modelopt/onnx/quantization/qdq_utils.py

@@ -1022,6 +1022,140 @@ def replace_zero_scale_with_smallest_nonzero(onnx_model: onnx.ModelProto) -> onn
     return onnx_model
 
 
+# =============================================================================
+# Column-major weight storage transformation for execution providers that need it
+# =============================================================================
+
+
+def _apply_transpose_perm_to_shape(shape, perm):
+    """Apply transpose permutation to a shape to get the output shape.
+
+    Args:
+        shape: Input shape as a list/tuple
+        perm: Permutation indices
+
+    Returns:
+        Transposed shape or None if inputs are None
+    """
+    if shape is None or perm is None:
+        return None
+    return [shape[i] for i in perm]
+
+
+def insert_transpose_nodes_for_column_major(graph: gs.Graph):
+    """Add a single Transpose node after each DequantizeLinear for column-major weights.
+
+    This implements the simple transformation: A @ B = A @ ((B^T)^T)
+    where B^T is stored in the DequantizeLinear node, and we add a Transpose
+    node after DQ to recover B before the MatMul.
+
+    Graph transformation:
+        Before: DQ(W) -> MatMul/Gemm
+        After:  DQ(W^T) -> Transpose -> W -> MatMul/Gemm
+
+    Args:
+        graph: ONNX GraphSurgeon graph to modify in-place
+    """
+    nodes_to_add = []
+    dq_nodes_processed = set()
+
+    for node in graph.nodes:
+        if node.op in ["MatMul", "Gemm"]:
+            # Check if second input (weight) is from DequantizeLinear
+            weight_input = node.inputs[1]
+            if not isinstance(weight_input, gs.Variable):
+                continue
+
+            # Find the producer of the weight input
+            producer_nodes = [n for n in graph.nodes if weight_input in n.outputs]
+            if not producer_nodes:
+                continue
+
+            producer_node = producer_nodes[0]
+            if producer_node.op != DEQUANTIZE_NODE_NAME:
+                continue
+
+            # Skip if we already processed this DQ node
+            if producer_node.name in dq_nodes_processed:
+                continue
+            dq_nodes_processed.add(producer_node.name)
+
+            # For Gemm nodes, check if transB is already set
+            if node.op == "Gemm":
+                trans_b = False
+                if hasattr(node, "attrs") and "transB" in node.attrs:
+                    trans_b = node.attrs["transB"] > 0
+                if trans_b:
+                    logger.debug(f"Gemm node {node.name} already has transB=1, skipping")
+                    continue
+
+            # Get weight shape and dtype from DQ output
+            # DQ outputs W^T (transposed), shape is [N, K] instead of [K, N]
+            weight_shape = weight_input.shape if hasattr(weight_input, "shape") else None
+            weight_dtype = weight_input.dtype if hasattr(weight_input, "dtype") else None
+
+            # Permutation for 2D weights: [1, 0] to transpose back
+            # The stored weight is B^T (transposed), we need to get B back
+            # For 2D [N, K] (stored as transposed): perm [1, 0] -> [K, N] (original)
+            perm = [1, 0]
+
+            # Compute the transposed shape (original weight shape)
+            transposed_weight_shape = _apply_transpose_perm_to_shape(weight_shape, perm)
+
+            # Create output variable for the transpose node
+            transpose_out = gs.Variable(
+                f"{producer_node.name}_transposed_back",
+                dtype=weight_dtype,
+                shape=transposed_weight_shape,
+            )
+
+            # Create transpose node: (B^T)^T = B
+            transpose_node = gs.Node(
+                op="Transpose",
+                name=f"{producer_node.name}_transpose_back",
+                inputs=[weight_input],
+                outputs=[transpose_out],
+                attrs={"perm": perm},
+            )
+
+            # Update MatMul/Gemm to use the transposed weight
+            node.inputs[1] = transpose_out
+
+            # Add transpose node to list
+            nodes_to_add.append(transpose_node)
+
+    # Add all new nodes to graph
+    if nodes_to_add:
+        graph.nodes.extend(nodes_to_add)
+        logger.info(f"Added {len(nodes_to_add)} transpose nodes for column-major optimization")
+
+    # Clean up and reorder graph
+    graph.cleanup().toposort()
+
+
+def apply_column_major_transformation(
+    gemm_weights_quantized: dict,
+    scales: dict,
+) -> None:
+    """Transpose quantized weights and scales in-place for column-major storage.
+
+    Note: After calling this function and inserting DQ nodes with axis=1,
+    you should call insert_transpose_nodes_for_column_major() on the graph.
+
+    Args:
+        gemm_weights_quantized: Dictionary mapping weight names to quantized weight arrays
+        scales: Dictionary mapping weight names to scale arrays
+    """
+    logger.info("Applying column-major storage optimization")
+
+    # Transpose weights and scales in-place
+    for name in list(gemm_weights_quantized.keys()):
+        gemm_weights_quantized[name] = gemm_weights_quantized[name].T
+
+    for name in list(scales.keys()):
+        scales[name] = scales[name].T
+
+
 def cast_initializer_to_dtype(
     node: onnx.NodeProto, dtype: str, initializer_map: dict[str, onnx.TensorProto]
 ):