tpu-inference 0.11.1.dev202511150811__py3-none-any.whl

tpu_inference/kernels/quantized_matmul/kernel.py

@@ -0,0 +1,395 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Quantized matmul kernel."""
+
+import functools
+
+import jax
+import jax.numpy as jnp
+from jax._src import dtypes
+from jax.experimental import pallas as pl
+from jax.experimental.pallas import tpu as pltpu
+
+from tpu_inference.kernels.quantized_matmul.tuned_block_sizes import (
+    TunedValue, get_device_vmem_limit, get_tuned_block_sizes)
+from tpu_inference.kernels.quantized_matmul.util import (get_kernel_name,
+                                                         next_multiple,
+                                                         unfold_args)
+
+
+def quantize_array(
+    x: jax.Array,  # [bs_block_size, in_block_size]
+    x_abs_max: jax.Array,  # [1, bs_block_size]
+    quant_dtype: jnp.dtype,
+):
+    is_float = jnp.issubdtype(quant_dtype, jnp.floating)
+    dtype_info = jnp.finfo(quant_dtype) if is_float else jnp.iinfo(quant_dtype)
+    dtype_max = float(dtype_info.max)
+
+    # TODO(kyuyeunk): Investigate performance gain from non xlu transpose.
+    scale = jnp.transpose(x_abs_max / dtype_max)
+    return (x / scale).astype(quant_dtype), scale.astype(jnp.float32)
+
+
+def get_vmem_limit(
+    n_batch: int,
+    n_out: int,
+    n_in: int,
+    batch_block_size: int,
+    out_block_size: int,
+    in_block_size: int,
+    x_dtype: jnp.dtype,
+    x_q_dtype: jnp.dtype,
+    w_q_dtype: jnp.dtype,
+    scale_dtype: jnp.dtype,
+    out_dtype: jnp.dtype,
+    acc_dtype: jnp.dtype,
+    save_acc: bool,
+    save_x_q: bool,
+    upper_limit_bytes: int,
+):
+    """Calculate VMEM limit for the kernel."""
+
+    # Calculate in/out VMEM size.
+    x_size = batch_block_size * in_block_size * dtypes.bit_width(x_dtype)
+    x_abs_max_size = batch_block_size * dtypes.bit_width(scale_dtype)
+    w_q_size = out_block_size * in_block_size * dtypes.bit_width(w_q_dtype)
+    w_scale_size = out_block_size * dtypes.bit_width(scale_dtype)
+    out_size = batch_block_size * out_block_size * dtypes.bit_width(out_dtype)
+
+    vmem_in_out = x_size + x_abs_max_size + w_q_size + w_scale_size + out_size
+    vmem_in_out *= 2  # Account for compute and vreg spills.
+
+    # Account for double buffering.
+    # Double buffering is used only if there are multiple blocks per in/out.
+    vmem_in_out += x_size if (n_batch > 1 or n_in > 1) else 0
+    vmem_in_out += x_abs_max_size if (n_batch > 1) else 0
+    vmem_in_out += w_q_size if (n_out > 1 or n_in > 1) else 0
+    vmem_in_out += w_scale_size if (n_out > 1) else 0
+    vmem_in_out += out_size if (n_batch > 1 or n_out > 1) else 0
+
+    # Calculate scratch VMEM size.
+    acc_size = batch_block_size * out_block_size * dtypes.bit_width(acc_dtype)
+    x_q_size = batch_block_size * in_block_size * dtypes.bit_width(x_q_dtype)
+    x_scale_size = batch_block_size * dtypes.bit_width(scale_dtype)
+
+    vmem_scratch = acc_size if save_acc else 0
+    vmem_scratch += x_q_size + x_scale_size if save_x_q else 0
+    vmem_scratch *= 2  # Account for compute and vreg spills.
+
+    # Add in/out and scratch VMEM size.
+    vmem_used = vmem_in_out + vmem_scratch
+    vmem_used_bytes = vmem_used // 8  # Convert bits to bytes.
+    # Specify upper limit. Defaults to 96MB.
+    vmem_limit_bytes = min(vmem_used_bytes, upper_limit_bytes)
+
+    return vmem_limit_bytes
+
+
+def validate_inputs(
+    x: jax.Array,
+    w_q: jax.Array,
+    w_scale: jax.Array,
+    x_abs_max: jax.Array,
+    x_q_dtype: jnp.dtype,
+    batch_block_size: int,
+    out_block_size: int,
+    in_block_size: int,
+):
+    """Verify inputs invoking the kernel."""
+
+    if x.dtype != x_q_dtype:
+        # If the input is quantized, then it should be the same subdtype as w_q
+        if jnp.issubdtype(x_q_dtype, jnp.integer) != jnp.issubdtype(
+                w_q.dtype, jnp.integer):
+            raise ValueError(
+                f'{x_q_dtype=} and {w_q.dtype=} must be the same int or float type.'
+            )
+
+    # Verify input shapes.
+    if x.shape[1] != w_q.shape[1]:
+        raise ValueError(f'{x.shape[1]=} must be equal to {w_q.shape[1]=}')
+    if w_q.shape[0] != w_scale.shape[1]:
+        raise ValueError(
+            f'{w_q.shape[0]=} must be equal to {w_scale.shape[1]=}')
+    if x_abs_max.shape != (1, x.shape[0]):
+        raise ValueError(
+            f'{x_abs_max.shape=} must be equal to (1, {x.shape[0]=})')
+    if x.shape[0] % batch_block_size != 0:
+        raise ValueError(
+            f'{x.shape[0]=} must be a multiple of {batch_block_size=}')
+    if w_q.shape[0] % out_block_size != 0:
+        raise ValueError(
+            f'{w_q.shape[0]=} must be a multiple of {out_block_size=}')
+    if x.shape[1] % in_block_size != 0:
+        raise ValueError(
+            f'{x.shape[1]=} must be a multiple of {in_block_size=}')
+
+
+def matmul_kernel(
+    x_ref: jax.Array,  # (batch_block_size, in_block_size)
+    w_q_ref: jax.Array,  # (out_block_size, in_block_size)
+    w_scale_ref: jax.Array,  # (1, out_block_size)
+    x_abs_max_ref: jax.Array,  # (1, batch_block_size)
+    out_ref: jax.Array,  # (batch_block_size, out_block_size)
+    acc_scratch: jax.Array,  # (batch_block_size, out_block_size)
+    x_q_scratch: jax.Array,  # (batch_block_size, in_block_size)
+    x_scale_scratch: jax.Array,  # (batch_block_size, 1)
+    *,
+    x_q_dtype: jnp.dtype,
+    save_acc: bool,
+    save_x_q: bool,
+):
+    out_idx, in_idx = pl.program_id(1), pl.program_id(2)
+    n_in = pl.num_programs(2)
+    x_ref_dtype = x_ref.dtype
+
+    quantize_activation = x_q_dtype != x_ref_dtype
+
+    # Initialize conditional logic.
+    if save_x_q:
+        assert quantize_activation
+        assert x_q_scratch is not None
+        assert x_scale_scratch is not None
+        quant = out_idx == 0
+    else:
+        assert x_q_scratch is None
+        assert x_scale_scratch is None
+        quant = quantize_activation
+
+    if save_acc:
+        assert acc_scratch is not None
+        is_first_step = in_idx == 0
+        is_last_step = in_idx == (n_in - 1)
+    else:
+        assert acc_scratch is None
+        is_first_step = True
+        is_last_step = True
+
+    acc_dtype = jnp.float32
+    if quantize_activation and jnp.issubdtype(w_q_ref.dtype, jnp.integer):
+        acc_dtype = jnp.int32
+
+    # Start of actual computation logic.
+    def matmul_body(quant: bool, is_first_step: bool, is_last_step: bool):
+        if quantize_activation:
+            if quant:
+                x_q_tmp, x_scale_tmp = quantize_array(
+                    x_ref[...],
+                    x_abs_max_ref[...],
+                    x_q_dtype,
+                )
+
+                if save_x_q:
+                    x_q_scratch[...] = x_q_tmp
+                    x_scale_scratch[...] = x_scale_tmp
+
+            else:
+                assert save_x_q
+                x_q_tmp = x_q_scratch[...]
+                if is_last_step:
+                    x_scale_tmp = x_scale_scratch[...]
+
+            acc = jax.lax.dot_general(
+                x_q_tmp,
+                w_q_ref[...],
+                (((1, ), (1, )), ((), ())),
+                preferred_element_type=acc_dtype,
+            )
+        else:
+            acc = jax.lax.dot_general(
+                x_ref[...],
+                w_q_ref[...],
+                (((1, ), (1, )), ((), ())),
+                preferred_element_type=acc_dtype,
+            )
+
+        if not is_first_step:
+            acc += acc_scratch[...]
+
+        if is_last_step:
+            acc *= w_scale_ref[...]
+            if quantize_activation:
+                # TODO(kyuyeunk): Investigate caching broadcast.
+                acc *= x_scale_tmp
+            out_ref[...] = acc.astype(x_ref_dtype)
+        else:
+            assert save_acc
+            acc_scratch[...] = acc
+
+    unfold_args((quant, is_first_step, is_last_step), (), matmul_body)
+
+
+@functools.partial(
+    jax.jit,
+    static_argnames=[
+        'x_q_dtype',
+        'tuned_value',
+    ],
+)
+def quantized_matmul_kernel(
+    x: jax.Array,  # [bs, n_in]
+    w_q: jax.Array,  # [n_out, n_in]
+    w_scale: jax.Array,  # [n_out]
+    w_zp: jax.Array | None = None,  # [n_out]
+    block_size: int | None = None,
+    x_q_dtype: jnp.dtype | None = None,
+    *,
+    tuned_value: TunedValue | None = None,
+) -> jax.Array:
+    """Quantized matmul kernel.
+
+  Args:
+    x: Input unquantized array.
+    w_q: Weight quantized array. [n_output_features, n_input_features]
+    w_scale: Weight quantization scale. [n_output_features]
+    w_zp: Weight zero point for asymmetric quantization.
+    block_size: Block size for subchannel quantization.
+    x_q_dtype: Quantization type of the input. If None or if the value is the
+      same as x.dtype, then no quantization is applied.
+    tuned_value: Kernel tuned values for optimal performance.
+
+  Returns:
+    Quantized matmul result.
+  """
+
+    if w_zp is not None:
+        raise NotImplementedError('zero_point is not supported.')
+    if block_size is not None:
+        raise NotImplementedError('block_size is not supported.')
+
+    if x_q_dtype is None:
+        x_q_dtype = x.dtype
+    quantize_activation = x_q_dtype != x.dtype
+
+    # Pallas kernel only has access to a single block of the input. Therefere,
+    # for per-token quantization, abs max has to be computed outside of the
+    # kernel.
+    x_abs_max = jnp.max(jnp.abs(x), axis=-1, keepdims=False)  # [bs]
+    # Pallas requires minormost dim to be a multiple of sublane size 128.
+    # Therefore, instead of using [bs, 1], we reshape this into [1, bs]
+    x_abs_max = jnp.expand_dims(x_abs_max, axis=0)  # [1, bs]
+    assert x_abs_max.shape == (1, x.shape[0])
+
+    orig_n_batch, orig_n_in = x.shape
+    orig_n_out, _ = w_q.shape
+
+    if tuned_value is None:
+        tuned_value = get_tuned_block_sizes(
+            n_batch=orig_n_batch,
+            n_out=orig_n_out,
+            n_in=orig_n_in,
+            x_q_dtype=jnp.dtype(x_q_dtype).name,
+            w_q_dtype=jnp.dtype(w_q.dtype).name,
+        )
+    batch_block_size = tuned_value.batch_block_size
+    out_block_size = tuned_value.out_block_size
+    in_block_size = tuned_value.in_block_size
+
+    # Pad the inputs to be multiple of block size.
+    padded_n_batch = next_multiple(orig_n_batch, batch_block_size)
+    if orig_n_batch < padded_n_batch:
+        x = jnp.pad(x, ((0, padded_n_batch - orig_n_batch), (0, 0)))
+        x_abs_max = jnp.pad(x_abs_max,
+                            ((0, 0), (0, padded_n_batch - orig_n_batch)))
+    padded_n_out = next_multiple(orig_n_out, out_block_size)
+    if orig_n_out < padded_n_out:
+        w_q = jnp.pad(w_q, ((0, padded_n_out - orig_n_out), (0, 0)))
+        w_scale = jnp.pad(w_scale, (0, padded_n_out - orig_n_out))
+    padded_n_in = next_multiple(orig_n_in, in_block_size)
+    if orig_n_in < padded_n_in:
+        x = jnp.pad(x, ((0, 0), (0, padded_n_in - orig_n_in)))
+        w_q = jnp.pad(w_q, ((0, 0), (0, padded_n_in - orig_n_in)))
+
+    if w_scale.dtype != jnp.float32:
+        w_scale = w_scale.astype(jnp.float32)
+    w_scale = jnp.expand_dims(w_scale, axis=0)  # [1, n_output_features]
+
+    n_batch = padded_n_batch // batch_block_size
+    n_out = padded_n_out // out_block_size
+    n_in = padded_n_in // in_block_size
+
+    save_acc = n_in > 1
+    # Remove redundant input quantization logic by caching quantized input. For
+    # best performance, only enable this behavior when single input block is
+    # used per batch.
+    save_x_q = quantize_activation and n_in == 1 and n_out > 1
+
+    acc_dtype = jnp.float32
+    if quantize_activation and jnp.issubdtype(w_q.dtype, jnp.integer):
+        acc_dtype = jnp.int32
+
+    vmem_limit_bytes = get_vmem_limit(
+        n_batch=n_batch,
+        n_out=n_out,
+        n_in=n_in,
+        batch_block_size=batch_block_size,
+        out_block_size=out_block_size,
+        in_block_size=in_block_size,
+        x_dtype=x.dtype,
+        x_q_dtype=x_q_dtype,
+        w_q_dtype=w_q.dtype,
+        scale_dtype=jnp.float32,
+        out_dtype=x.dtype,
+        acc_dtype=acc_dtype,
+        save_acc=save_acc,
+        save_x_q=save_x_q,
+        upper_limit_bytes=get_device_vmem_limit(),
+    )
+
+    kernel = pl.pallas_call(
+        functools.partial(
+            matmul_kernel,
+            x_q_dtype=x_q_dtype,
+            save_acc=save_acc,
+            save_x_q=save_x_q,
+        ),
+        grid_spec=pltpu.PrefetchScalarGridSpec(
+            num_scalar_prefetch=0,
+            in_specs=[
+                pl.BlockSpec((batch_block_size, in_block_size), lambda b, o, i:
+                             (b, i)),  # x
+                pl.BlockSpec((out_block_size, in_block_size), lambda b, o, i:
+                             (o, i)),  # w_q
+                pl.BlockSpec((1, out_block_size), lambda b, o, i:
+                             (0, o)),  # w_scale
+                pl.BlockSpec((1, batch_block_size), lambda b, o, i:
+                             (0, b)),  # x_abs_max
+            ],
+            out_specs=pl.BlockSpec((batch_block_size, out_block_size),
+                                   lambda b, o, i: (b, o)),
+            scratch_shapes=[
+                pltpu.VMEM((batch_block_size, out_block_size), acc_dtype)
+                if save_acc else None,  # acc_scratch
+                pltpu.VMEM((batch_block_size, in_block_size), x_q_dtype)
+                if save_x_q else None,  # x_q_scratch
+                pltpu.VMEM(
+                    (batch_block_size,
+                     1), jnp.float32) if save_x_q else None,  # x_scale_scratch
+            ],
+            grid=(n_batch, n_out, n_in),
+        ),
+        out_shape=jax.ShapeDtypeStruct((padded_n_batch, padded_n_out),
+                                       x.dtype),
+        compiler_params=pltpu.CompilerParams(
+            dimension_semantics=('parallel', 'arbitrary', 'arbitrary'),
+            vmem_limit_bytes=vmem_limit_bytes,
+        ),
+    )
+
+    validate_inputs(
+        x=x,
+        w_q=w_q,
+        w_scale=w_scale,
+        x_abs_max=x_abs_max,
+        x_q_dtype=x_q_dtype,
+        batch_block_size=batch_block_size,
+        out_block_size=out_block_size,
+        in_block_size=in_block_size,
+    )
+
+    # The named_scope is used for autotune.
+    kernel_name = get_kernel_name(tuned_value)
+    with jax.named_scope(kernel_name):
+        out = kernel(x, w_q, w_scale, x_abs_max)
+
+    return out[:orig_n_batch, :orig_n_out]