draft: Add per block MSE for NVFP4 and INT4

modelopt/torch/quantization/calib/mse.py

@@ -79,8 +79,17 @@ def collect(self, x: torch.Tensor):
         x = x.detach().to(dtype=torch.float32)
 
         device = x.device
-        multipliers = torch.linspace(
-            self._start_multiplier, self._stop_multiplier, steps=self._num_steps, device=device
+        # Split steps between _start_multiplier to 1.0 and 1.0 to _stop_multiplier
+        # to ensure balanced exploration on both sides of the original amax (1.0)
+        steps_first_half = self._num_steps // 2 + 1  # Include 1.0
+        steps_second_half = self._num_steps - self._num_steps // 2  # For second range
+        multipliers = torch.cat(
+            [
+                torch.linspace(self._start_multiplier, 1.0, steps=steps_first_half, device=device),
+                torch.linspace(1.0, self._stop_multiplier, steps=steps_second_half, device=device)[
+                    1:
+                ],  # Skip duplicate 1.0
+            ]
         )
 
         # Get reduce axis for per-channel quantization

modelopt/torch/quantization/model_calib.py

@@ -32,6 +32,7 @@
 from .calib import MseCalibrator
 from .conversion import create_and_replace_svdquant_linear_on_the_fly, set_quantizer_by_cfg_context
 from .nn import QuantModule, SequentialQuantizer, TensorQuantizer
+from .tensor_quant import scaled_e4m3_impl
 from .utils import (
     disable_calib,
     enable_fake_quant,
@@ -41,6 +42,7 @@
     is_quantized_linear,
     is_quantized_row_parallel_linear,
     quantizer_attr_names,
+    reduce_amax,
     weight_attr_names,
 )
 
@@ -216,14 +218,18 @@ def mse_calibrate(
     max_calibrate(model, forward_loop, distributed_sync)
 
     # Step 2: Replace calibrators with MseCalibrator for enabled quantizers
+    # and identify weight quantizers
+    weight_quantizers = []
+    seen_modules = set()
+
     for name, module in model.named_modules():
         if isinstance(module, TensorQuantizer) and not module._disabled:
             # Static block quantization is not supported by MseCalibrator
-            if module.is_static_block_quant:
-                raise ValueError(
-                    f"MSE calibration does not support static block quantization. "
-                    f"Found static block quantization at {name}."
-                )
+            # if module.is_static_block_quant:
+            #     raise ValueError(
+            #         f"MSE calibration does not support static block quantization. "
+            #         f"Found static block quantization at {name}."
+            #     )
             if module._calibrator is not None and not module._dynamic and hasattr(module, "_amax"):
                 # Get the initial amax from max calibration
                 initial_amax = module._amax.clone().detach()
@@ -237,7 +243,20 @@ def quant_func(x, amax, quantizer=module):
                         disable_calib(quantizer),
                         enable_fake_quant(quantizer),
                     ):
+                        quantizer._keep_shape = True
                         xq = quantizer(x)
+                        quantizer._keep_shape = False
+
+                        # FP8 quantization of NVFP4 static per-block scales
+                        if (
+                            quantizer.is_static_block_quant
+                            and quantizer._num_bits == (2, 1)
+                            and quantizer._block_sizes.get("scale_bits") == (4, 3)
+                        ):
+                            weight_amax = reduce_amax(
+                                x, axis=None, keepdims=False, squeeze_scalar=True
+                            )
+                            quantizer._amax = scaled_e4m3_impl(amax, weight_amax)
 
                     if original_amax is not None:
                         quantizer._amax = original_amax
@@ -256,14 +275,48 @@ def quant_func(x, amax, quantizer=module):
                     quant_func=quant_func,
                 )
 
-    # Step 3: Collect data with MSE calibrators
+    # Identify weight quantizers by checking if they have corresponding weight parameters
+    for name, parent_module in model.named_modules():
+        if parent_module in seen_modules:
+            continue
+        for weight_name in weight_attr_names(parent_module):
+            weight_quantizer_name = quantizer_attr_names(weight_name).weight_quantizer
+            weight_quantizer = getattr(parent_module, weight_quantizer_name, None)
+            if isinstance(weight_quantizer, TensorQuantizer) and not weight_quantizer._disabled:
+                if weight_quantizer._calibrator is not None:
+                    weight_quantizers.append((parent_module, weight_name, weight_quantizer))
+        seen_modules.add(parent_module)
+
+    # Step 3: Calibrate weight quantizers once with MSE calibration
+    # This ensures weights are only calibrated once, not during every forward pass
+    for parent_module, weight_name, weight_quantizer in weight_quantizers:
+        # Enable calibration mode for the weight quantizer
+        weight_quantizer.disable_quant()
+        weight_quantizer.enable_calib()
+
+        with enable_weight_access_and_writeback(parent_module, model):
+            weight = getattr(parent_module, weight_name)
+            weight_quantizer(weight)
+
+    # Step 4: Disable weight quantizers during forward loop
+    for _, _, weight_quantizer in weight_quantizers:
+        weight_quantizer.disable()
+
+    # Step 5: Collect data with MSE calibrators for activation quantizers only
     enable_stats_collection(model)
     if forward_loop is None:
-        weight_only_quantize(model)
+        # If no forward loop, nothing else to do since weights are already calibrated
+        pass
     else:
+        # Run forward loop - only activation quantizers will collect data
         forward_loop(model)
 
-    # Step 4: Compute optimal amax and load it
+    # Step 6: Re-enable weight quantizers before finalizing calibration
+    # This ensures finish_stats_collection processes them correctly
+    for _, _, weight_quantizer in weight_quantizers:
+        weight_quantizer.enable()
+
+    # Step 7: Compute optimal amax and load it for all quantizers (weights + activations)
     finish_stats_collection(model, method="mse")
 
     # TODO: Sync amax across distributed processes

modelopt/torch/quantization/nn/modules/tensor_quantizer.py

@@ -128,6 +128,7 @@ def __init__(
         self._enable_pre_quant_scale = True
         self._dequantize = False
         self._input_dtype = None
+        self._keep_shape = False
 
         # Lazy initialize the bias calibrator for KV cache quantization
         self._bias_calibrator = None
@@ -653,6 +654,14 @@ def _fake_quantize(self, inputs):
                 getattr(self, "_onnx_quantizer_type", None),
                 self._pass_through_bwd,
             )
+        elif self._num_bits == (2, 1) and self.is_static_block_quant:
+            from modelopt.torch.quantization.triton.fp4_kernel import (
+                launch_static_blockwise_fp4_fake_quant,
+            )
+
+            outputs = launch_static_blockwise_fp4_fake_quant(
+                inputs, amax / 6.0, out_dtype=inputs.dtype
+            )
         elif isinstance(self._num_bits, tuple):
             # Float-point quantization, e.g., FP8
             E, M = self._num_bits  # noqa: N806
@@ -783,11 +792,11 @@ def _process_for_blockquant(self, inputs: torch.Tensor):
         if hasattr(self, "_padding"):
             inputs = F.pad(inputs, self._padding, "constant", 0)
 
-        if inputs.shape != self._original_shape:
-            raise ValueError(
-                f"Input shape has changed from {self._original_shape} to {inputs.shape}."
-                " Block-quantization requires a fixed input shape."
-            )
+        # if inputs.shape != self._original_shape:
+        #     print(
+        #         f"Input shape has changed from {self._original_shape} to {inputs.shape}."
+        #         " Block-quantization requires a fixed input shape."
+        #     )
         inputs = inputs.reshape(self._block_reshape_size)
         return inputs
 
@@ -941,7 +950,7 @@ def forward(self, inputs):
                     "This case should have been handled."
                 )
 
-        if self.is_static_block_quant:
+        if self.is_static_block_quant and not self._keep_shape:
             outputs = self._reset_to_original_shape(outputs)
 
         return outputs

modelopt/torch/quantization/triton/fp4_kernel.py

@@ -345,3 +345,157 @@ def fp4_dequantize(
     )
 
     return output
+
+
+@triton.jit
+def static_blockwise_fp4_fake_quant_kernel(
+    x_ptr,  # [NUM_FP4_BLOCKS * BLOCK_SIZE]
+    y_ptr,  # [NUM_FP4_BLOCKS * BLOCK_SIZE]
+    scale_ptr,  # [NUM_FP4_BLOCKS]
+    NUM_FP4_BLOCKS,
+    BLOCK_SIZE: tl.constexpr,
+    OUT_DTYPE: tl.constexpr,
+):
+    pid = tl.program_id(axis=0)
+    if pid >= NUM_FP4_BLOCKS:
+        return
+
+    block_offset = pid * BLOCK_SIZE
+    idx = block_offset + tl.arange(0, BLOCK_SIZE)
+
+    scale = tl.load(scale_ptr + pid).to(tl.float32)
+
+    x = tl.load(x_ptr + idx).to(tl.float32)
+
+    x_abs = tl.abs(x)
+    scale_safe = tl.where(scale >= 1e-5, scale, 1.0)
+    abs_scaled = x_abs / scale_safe
+
+    # FP4 values: 0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0
+    q_val = tl.where(
+        abs_scaled <= 0.25,
+        0.0,
+        tl.where(
+            abs_scaled < 0.75,
+            0.5,
+            tl.where(
+                abs_scaled <= 1.25,
+                1.0,
+                tl.where(
+                    abs_scaled < 1.75,
+                    1.5,
+                    tl.where(
+                        abs_scaled <= 2.5,
+                        2.0,
+                        tl.where(
+                            abs_scaled < 3.5,
+                            3.0,
+                            tl.where(abs_scaled <= 5.0, 4.0, 6.0),
+                        ),
+                    ),
+                ),
+            ),
+        ),
+    )
+
+    x_rescaled = q_val * scale_safe
+    x_dequant = tl.where(x >= 0, x_rescaled, -x_rescaled)
+
+    tl.store(y_ptr + idx, x_dequant.to(OUT_DTYPE))
+
+
+def launch_static_blockwise_fp4_fake_quant(
+    x: torch.Tensor,
+    scale: torch.Tensor,
+    out_dtype: torch.dtype = torch.float16,
+):
+    """Launch Triton kernel for blockwise FP4 fake quantization.
+
+    x: [NUM_FP4_BLOCKS, BLOCK_SIZE] on CUDA.
+    scale: [NUM_FP4_BLOCKS] or [NUM_FP4_BLOCKS, 1] on CUDA.
+    """
+    assert x.ndim == 2
+    NUM_FP4_BLOCKS, BLOCK_SIZE = x.shape
+
+    x_flat = x.contiguous().view(-1)
+    y_flat = torch.empty_like(x_flat, dtype=out_dtype)
+    scale_flat = scale.view(NUM_FP4_BLOCKS).contiguous()
+
+    tl_out_dtype = _torch_dtype_to_tl(out_dtype)
+
+    grid = (NUM_FP4_BLOCKS,)
+
+    # Ensure we're running on the correct CUDA device
+    with torch.cuda.device(x.device):
+        static_blockwise_fp4_fake_quant_kernel[grid](
+            x_flat,
+            y_flat,
+            scale_flat,
+            NUM_FP4_BLOCKS,
+            BLOCK_SIZE,
+            OUT_DTYPE=tl_out_dtype,
+        )
+
+    return y_flat.view_as(x)
+
+
+def blockwise_fp4_fake_quant_reference(
+    x: torch.Tensor,
+    scale: torch.Tensor,
+    out_dtype: torch.dtype = torch.bfloat16,
+) -> torch.Tensor:
+    """Reference implementation of blockwise FP4 fake quantization.
+
+    x: [NUM_FP4_BLOCKS, BLOCK_SIZE].
+    scale: [NUM_FP4_BLOCKS] or [NUM_FP4_BLOCKS, 1].
+
+    Uses FP4 quantization levels: 0.0, 0.5, 1.0, 1.5, 2.0, 3.0, 4.0, 6.0.
+    """
+    assert x.ndim == 2
+    num_blocks, block_size = x.shape
+
+    if scale.ndim == 1:
+        scale = scale.view(num_blocks, 1)
+    assert scale.shape == (num_blocks, 1)
+
+    x_f = x.to(torch.float32)
+    s_f = scale.to(torch.float32)
+
+    s_f = torch.where(s_f >= 1e-5, s_f, torch.ones_like(s_f))
+
+    x_abs = torch.abs(x_f)
+    abs_scaled = x_abs / s_f
+
+    q_val = torch.where(
+        abs_scaled <= 0.25,
+        torch.zeros_like(abs_scaled),
+        torch.where(
+            abs_scaled < 0.75,
+            torch.full_like(abs_scaled, 0.5),
+            torch.where(
+                abs_scaled <= 1.25,
+                torch.ones_like(abs_scaled),
+                torch.where(
+                    abs_scaled < 1.75,
+                    torch.full_like(abs_scaled, 1.5),
+                    torch.where(
+                        abs_scaled <= 2.5,
+                        torch.full_like(abs_scaled, 2.0),
+                        torch.where(
+                            abs_scaled < 3.5,
+                            torch.full_like(abs_scaled, 3.0),
+                            torch.where(
+                                abs_scaled <= 5.0,
+                                torch.full_like(abs_scaled, 4.0),
+                                torch.full_like(abs_scaled, 6.0),
+                            ),
+                        ),
+                    ),
+                ),
+            ),
+        ),
+    )
+
+    x_rescaled = q_val * s_f
+    x_dequant = torch.where(x_f >= 0, x_rescaled, -x_rescaled)
+    return x_dequant.to(out_dtype)