pre big pr cleanup

src/chop/nn/quantizers/__init__.py

@@ -9,7 +9,6 @@
 from .quantizers_for_hw import (
     integer_quantizer_for_hw,
     integer_floor_quantizer_for_hw,
-    mxint_quantizer_for_hw,
 )
 from .mxint import mxint_quantizer
 

src/chop/nn/quantizers/quantizers_for_hw.py

@@ -35,64 +35,3 @@ def integer_floor_quantizer_for_hw(x: Tensor, width: int, frac_width: int):
     fixed_point_value = fixed_point_value.to(torch.int)
     fixed_point_value = fixed_point_value % (2**width)
     return fixed_point_value
-
-
-def mxint_quantizer_for_hw(
-    x: Tensor,
-    width: int,
-    exponent_width: int,
-    block_size: list[int],
-    floor: bool = False,
-):
-    """
-    - Convert IEEE FP32/64 to Microscaling Interger (MXINT), where an exponent is shared over all elements in a block.
-    - https://arxiv.org/pdf/2310.10537.pdf
-    - https://www.opencompute.org/documents/ocp-microscaling-formats-mx-v1-0-spec-final-pdf
-
-
-    ---
-    - forward: convert IEEE FP32/64 to MXINT
-    - backward: STE
-
-    ---
-    - `width`: The number of mantissa bits + 1 (the sign bit)
-    - `exponent_width`: the number of exponent bits
-    - `block_size`: a list of integers where each integer is the block size on that dimension. See function `block`.
-    """
-
-    blocked_x, per_block_max, padded_x_shape, block_shape = block(
-        x,
-        block_shape=block_size,
-    )
-
-    if torch.all(per_block_max == 0):
-        per_block_max = torch.ones_like(per_block_max)
-    else:
-        per_block_max[per_block_max == 0] = per_block_max[per_block_max != 0].min()
-
-    if torch.all(per_block_max == 0):
-        per_block_max = torch.ones_like(per_block_max)
-    else:
-        per_block_max[per_block_max == 0] = per_block_max[per_block_max != 0].min()
-
-    exponent_bias = 2 ** (exponent_width - 1) - 1
-
-    per_block_expontent = my_floor(torch.log2(per_block_max)) + exponent_bias
-    per_block_expontent = my_clamp(per_block_expontent, 0, 2**exponent_width - 1)
-
-    element_max = 2 ** (width - 1) - 1
-    shift = 2 ** (width - 2)
-
-    scaled_value = shift * blocked_x / 2 ** (per_block_expontent - exponent_bias)
-
-    if floor:
-        quantized_value = my_floor(scaled_value)
-    else:
-        quantized_value = my_round(scaled_value)
-
-    quantized_value = my_clamp(quantized_value, -element_max, element_max)
-
-    return quantized_value, per_block_expontent
-
-
-# sw_quantizer_to_hw_quantizer = {integer_quantizer: integer_quantizer_for_hw}

src/chop/passes/graph/transforms/verilog/emit_bram.py

@@ -17,7 +17,6 @@
 from chop.nn.quantizers import (
     integer_quantizer_for_hw,
     integer_floor_quantizer_for_hw,
-    mxint_quantizer_for_hw,
 )
 
 logger = logging.getLogger(__name__)
@@ -307,8 +306,6 @@ def emit_parameters_in_dat_internal(node, param_name, file_name):
             else:
                 base_quantizer = integer_quantizer_for_hw
 
-            scale = 2**frac_width
-            thresh = 2**width
             data_buff = ""
             for i in range(0, out_depth):
                 line_buff = ""