Build uploaded using `kernels`.

build/torch27-cxx11-cu128-aarch64-linux/causal_conv1d/{_causal_conv1d_5b5a6b7.abi3.so → _causal_conv1d_306ae84.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:f53abe29129caaa23e5e9d730b873cd426835679a362b8927dce9b52880a3bc0
-size 102592040
+oid sha256:7ec3c41413afbb69d499eae6a432fa9d41a580e7b1c6ee83d09e8dab51f91803
+size 90795560

build/torch27-cxx11-cu128-aarch64-linux/causal_conv1d/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _causal_conv1d_5b5a6b7
-ops = torch.ops._causal_conv1d_5b5a6b7
+from . import _causal_conv1d_306ae84
+ops = torch.ops._causal_conv1d_306ae84
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_causal_conv1d_5b5a6b7::{op_name}"
+    return f"_causal_conv1d_306ae84::{op_name}"

build/torch28-cxx11-cu129-aarch64-linux/causal_conv1d/{_causal_conv1d_5b5a6b7.abi3.so → _causal_conv1d_306ae84.abi3.so}

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:6896f3385fb9d876c29eec9e015666a760ba6e6ba3f3b4ae2c3c9a800fde4be2
-size 110261680
+oid sha256:4c3535b795cbc5baf363b0cd8636649153b017599c9992ea6c08aa4ab23ceae0
+size 97678768

build/torch28-cxx11-cu129-aarch64-linux/causal_conv1d/_ops.py

@@ -1,9 +1,9 @@
 import torch
-from . import _causal_conv1d_5b5a6b7
-ops = torch.ops._causal_conv1d_5b5a6b7
+from . import _causal_conv1d_306ae84
+ops = torch.ops._causal_conv1d_306ae84
 
 def add_op_namespace_prefix(op_name: str):
     """
     Prefix op by namespace.
     """
-    return f"_causal_conv1d_5b5a6b7::{op_name}"
+    return f"_causal_conv1d_306ae84::{op_name}"

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/__init__.py

@@ -0,0 +1,4 @@
+from .causal_conv1d_interface import causal_conv1d_fn, causal_conv1d_update
+from .causal_conv1d_varlen import causal_conv1d_varlen_states
+
+__all__ = ["causal_conv1d_fn", "causal_conv1d_update", "causal_conv1d_varlen_states"]

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/_causal_conv1d_306ae84.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d069be88c48d9a42dd8cceeabe1043f99f53196ada6760a85c2fc6b3bee22cfe
+size 64340360

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _causal_conv1d_306ae84
+ops = torch.ops._causal_conv1d_306ae84
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_causal_conv1d_306ae84::{op_name}"

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/causal_conv1d_interface.py

@@ -0,0 +1,242 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+import torch.nn.functional as F
+
+from .cpp_functions import causal_conv1d_fwd_function, causal_conv1d_bwd_function, causal_conv1d_update_function
+
+
+class CausalConv1dFn(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x,
+        weight,
+        bias=None,
+        seq_idx=None,
+        initial_states=None,
+        return_final_states=False,
+        final_states_out=None,
+        activation=None,
+    ):
+        if activation not in [None, "silu", "swish"]:
+            raise NotImplementedError("activation must be None, silu, or swish")
+        if x.stride(2) != 1 and x.stride(1) != 1:
+            x = x.contiguous()
+        bias = bias.contiguous() if bias is not None else None
+        if seq_idx is not None:
+            assert (
+                initial_states is None
+            ), "initial_states must be None if seq_idx is not None"
+            assert (
+                not return_final_states
+            ), "If seq_idx is not None, we don't return final_states_out"
+        seq_idx = seq_idx.contiguous() if seq_idx is not None else None
+        if initial_states is not None and (
+            initial_states.stride(2) != 1 and initial_states.stride(1) != 1
+        ):
+            initial_states = initial_states.contiguous()
+        if return_final_states:
+            assert (
+                x.stride(1) == 1
+            ), "Only channel-last layout support returning final_states_out"
+            if final_states_out is not None:
+                assert (
+                    final_states_out.stride(2) == 1 or final_states_out.stride(1) == 1
+                )
+            else:
+                batch, dim, seqlen = x.shape
+                width = weight.shape[1]
+                final_states_out = torch.empty(
+                    batch, width - 1, dim, device=x.device, dtype=x.dtype
+                ).transpose(1, 2)
+        else:
+            final_states_out = None
+        ctx.activation = activation in ["silu", "swish"]
+        out = causal_conv1d_fwd_function(
+            x, weight, bias, seq_idx, initial_states, final_states_out, ctx.activation
+        )
+        ctx.save_for_backward(x, weight, bias, seq_idx, initial_states)
+        ctx.return_final_states = return_final_states
+        ctx.return_dinitial_states = (
+            initial_states is not None and initial_states.requires_grad
+        )
+        return out if not return_final_states else (out, final_states_out)
+
+    @staticmethod
+    def backward(ctx, dout, *args):
+        x, weight, bias, seq_idx, initial_states = ctx.saved_tensors
+        dfinal_states = args[0] if ctx.return_final_states else None
+        if dout.stride(2) != 1 and dout.stride(1) != 1:
+            dout = dout.contiguous()
+        # The kernel supports passing in a pre-allocated dx (e.g., in case we want to fuse the
+        # backward of conv1d with the backward of chunk).
+        # Here we just pass in None and dx will be allocated in the C++ code.
+        dx, dweight, dbias, dinitial_states = causal_conv1d_bwd_function(
+            x,
+            weight,
+            bias,
+            dout,
+            seq_idx,
+            initial_states,
+            dfinal_states,
+            None,
+            ctx.return_dinitial_states,
+            ctx.activation,
+        )
+        return (
+            dx,
+            dweight,
+            dbias if bias is not None else None,
+            None,
+            dinitial_states if initial_states is not None else None,
+            None,
+            None,
+            None,
+        )
+
+
+def causal_conv1d_fn(
+    x,
+    weight,
+    bias=None,
+    seq_idx=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    seq_idx: (batch, seqlen)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1), to be written to
+    activation: either None or "silu" or "swish"
+
+    out: (batch, dim, seqlen)
+    """
+    return CausalConv1dFn.apply(
+        x,
+        weight,
+        bias,
+        seq_idx,
+        initial_states,
+        return_final_states,
+        final_states_out,
+        activation,
+    )
+
+
+def causal_conv1d_ref(
+    x,
+    weight,
+    bias=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1)
+
+    out: (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    x = x.to(weight.dtype)
+    seqlen = x.shape[-1]
+    dim, width = weight.shape
+    if initial_states is None:
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=width - 1, groups=dim)
+    else:
+        x = torch.cat([initial_states, x], dim=-1)
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=0, groups=dim)
+    out = out[..., :seqlen]
+    if return_final_states:
+        final_states = F.pad(x, (width - 1 - x.shape[-1], 0)).to(
+            dtype_in
+        )  # (batch, dim, width - 1)
+        if final_states_out is not None:
+            final_states_out.copy_(final_states)
+        else:
+            final_states_out = final_states
+    out = (out if activation is None else F.silu(out)).to(dtype=dtype_in)
+    return out if not return_final_states else (out, final_states_out)
+
+
+def causal_conv1d_update(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None, conv_state_indices=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len.
+    conv_state_indices: (batch,), dtype int32
+        If None, the conv_state is a larger tensor along the batch dim, 
+        and we are selecting the batch coords specified by conv_state_indices.
+        Useful for a continuous batching scenario.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    activation = activation in ["silu", "swish"]
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    out = causal_conv1d_update_function(
+        x, conv_state, weight, bias, activation, cache_seqlens, conv_state_indices
+    )
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return out
+
+
+def causal_conv1d_update_ref(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len before performing the convolution.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    batch, dim, seqlen = x.shape
+    width = weight.shape[1]
+    state_len = conv_state.shape[-1]
+    assert conv_state.shape == (batch, dim, state_len)
+    assert weight.shape == (dim, width)
+    if cache_seqlens is None:
+        x_new = torch.cat([conv_state, x], dim=-1).to(weight.dtype)  # (batch, dim, state_len + seqlen)
+        conv_state.copy_(x_new[:, :, -state_len:])
+    else:
+        width_idx = torch.arange(-(width - 1), 0, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        width_idx = torch.remainder(width_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        x_new = torch.cat([conv_state.gather(2, width_idx), x], dim=-1).to(weight.dtype)
+        copy_idx = torch.arange(seqlen, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        copy_idx = torch.remainder(copy_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        conv_state.scatter_(2, copy_idx, x)
+    out = F.conv1d(x_new, weight.unsqueeze(1), bias, padding=0, groups=dim)[:, :, -seqlen:]
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return (out if activation is None else F.silu(out)).to(dtype=dtype_in)

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/causal_conv1d_varlen.py

@@ -0,0 +1,86 @@
+import torch
+from torch import Tensor
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _causal_conv1d_varlen_states(
+    X,
+    CU_SEQLENS,
+    STATES,
+    state_len,
+    dim,
+    stride_x_seqlen, stride_x_dim,
+    stride_states_batch, stride_states_seqlen, stride_states_dim,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr
+):
+    batch_idx = tl.program_id(2)
+    STATES += batch_idx * stride_states_batch
+    end_idx = tl.load(CU_SEQLENS + batch_idx + 1)
+    start_idx = tl.maximum(tl.load(CU_SEQLENS + batch_idx), end_idx - state_len)
+    rows = end_idx - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    cols = tl.program_id(0) * BLOCK_N + tl.arange(0, BLOCK_N)
+    x = tl.load(X + rows[:, None] * stride_x_seqlen + cols[None, :] * stride_x_dim,
+                mask=(rows[:, None] >= start_idx) & (cols[None, :] < dim),
+                other=0)
+    rows_states = state_len - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    tl.store(STATES + rows_states[:, None] * stride_states_seqlen + cols[None, :] * stride_states_dim,
+             x,
+             mask=(rows_states[:, None] >= 0) & (cols[None, :] < dim))
+
+
+def causal_conv1d_varlen_states(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.empty(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    BLOCK_M = min(triton.next_power_of_2(state_len), 16)
+    BLOCK_N = min(triton.next_power_of_2(dim), 256)
+    grid = (triton.cdiv(dim, BLOCK_N), triton.cdiv(state_len, BLOCK_M), batch)
+    with torch.cuda.device(x.device.index):
+        _causal_conv1d_varlen_states[grid](
+            x,
+            cu_seqlens,
+            states,
+            state_len,
+            dim,
+            x.stride(0), x.stride(1),
+            states.stride(0), states.stride(2), states.stride(1),
+            BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N
+        )
+    return states
+
+
+def causal_conv1d_varlen_states_ref(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.zeros(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    for i in range(batch):
+        end_idx = cu_seqlens[i + 1]
+        start_idx = torch.maximum(cu_seqlens[i], end_idx - state_len)
+        states[i, :, -(end_idx - start_idx):] = x[start_idx:end_idx].T
+    return states

build/torch29-cxx11-cu126-aarch64-linux/causal_conv1d/cpp_functions.py

@@ -0,0 +1,96 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+
+from ._ops import ops
+
+def causal_conv1d_fwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    final_states_out: torch.Tensor | None,
+    silu_activation: bool,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_fwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        out=out,
+        final_states_out=final_states_out,
+        silu_activation=silu_activation,
+    )
+    return out
+
+
+def causal_conv1d_bwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    dout: torch.Tensor,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    dfinal_states: torch.Tensor | None,
+    dx: torch.Tensor | None,
+    return_dinitial_states: torch.Tensor,
+    silu_activation: bool,
+) -> tuple[torch.Tensor | None]:
+    batch_size, dim = x.size()[:2]
+    width = weight.size(-1)
+
+    if dx is None:
+        dx = torch.empty_like(x)
+    dweight = torch.zeros_like(weight, dtype=torch.float32)
+    dbias = None
+    if bias is not None:
+        dbias = torch.zeros_like(bias, dtype=torch.float32)
+    dinitial_states = None
+    if return_dinitial_states:
+        dinitial_states = torch.empty(batch_size, width - 1, dim, device=x.device, dtype=x.dtype).transpose(1, 2)
+
+    ops.causal_conv1d_bwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        dout=dout,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        dfinal_states=dfinal_states,
+        dx=dx,
+        dweight=dweight,
+        dbias=dbias,
+        dinitial_states=dinitial_states,
+        silu_activation=silu_activation,
+    )
+
+    dweight = dweight.type_as(weight)
+    if dbias is not None:
+        dbias = dbias.type_as(bias)
+    return dx, dweight, dbias, dinitial_states
+
+
+def causal_conv1d_update_function(
+    x: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    silu_activation: bool,
+    cache_seqlens: torch.Tensor | None,
+    conv_state_indices: torch.Tensor | None,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_update(
+        x=x,
+        conv_state=conv_state,
+        weight=weight,
+        bias=bias,
+        out=out,
+        silu_activation=silu_activation,
+        cache_seqlens=cache_seqlens,
+        conv_state_indices=conv_state_indices,
+    )
+    return out

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/__init__.py

@@ -0,0 +1,4 @@
+from .causal_conv1d_interface import causal_conv1d_fn, causal_conv1d_update
+from .causal_conv1d_varlen import causal_conv1d_varlen_states
+
+__all__ = ["causal_conv1d_fn", "causal_conv1d_update", "causal_conv1d_varlen_states"]

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/_causal_conv1d_306ae84.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b252d8148e4fdb48015fbf2cfc3a5c2683c414066b7b825d32a9f5578e3869d6
+size 90795712

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _causal_conv1d_306ae84
+ops = torch.ops._causal_conv1d_306ae84
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_causal_conv1d_306ae84::{op_name}"

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/causal_conv1d_interface.py

@@ -0,0 +1,242 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+import torch.nn.functional as F
+
+from .cpp_functions import causal_conv1d_fwd_function, causal_conv1d_bwd_function, causal_conv1d_update_function
+
+
+class CausalConv1dFn(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x,
+        weight,
+        bias=None,
+        seq_idx=None,
+        initial_states=None,
+        return_final_states=False,
+        final_states_out=None,
+        activation=None,
+    ):
+        if activation not in [None, "silu", "swish"]:
+            raise NotImplementedError("activation must be None, silu, or swish")
+        if x.stride(2) != 1 and x.stride(1) != 1:
+            x = x.contiguous()
+        bias = bias.contiguous() if bias is not None else None
+        if seq_idx is not None:
+            assert (
+                initial_states is None
+            ), "initial_states must be None if seq_idx is not None"
+            assert (
+                not return_final_states
+            ), "If seq_idx is not None, we don't return final_states_out"
+        seq_idx = seq_idx.contiguous() if seq_idx is not None else None
+        if initial_states is not None and (
+            initial_states.stride(2) != 1 and initial_states.stride(1) != 1
+        ):
+            initial_states = initial_states.contiguous()
+        if return_final_states:
+            assert (
+                x.stride(1) == 1
+            ), "Only channel-last layout support returning final_states_out"
+            if final_states_out is not None:
+                assert (
+                    final_states_out.stride(2) == 1 or final_states_out.stride(1) == 1
+                )
+            else:
+                batch, dim, seqlen = x.shape
+                width = weight.shape[1]
+                final_states_out = torch.empty(
+                    batch, width - 1, dim, device=x.device, dtype=x.dtype
+                ).transpose(1, 2)
+        else:
+            final_states_out = None
+        ctx.activation = activation in ["silu", "swish"]
+        out = causal_conv1d_fwd_function(
+            x, weight, bias, seq_idx, initial_states, final_states_out, ctx.activation
+        )
+        ctx.save_for_backward(x, weight, bias, seq_idx, initial_states)
+        ctx.return_final_states = return_final_states
+        ctx.return_dinitial_states = (
+            initial_states is not None and initial_states.requires_grad
+        )
+        return out if not return_final_states else (out, final_states_out)
+
+    @staticmethod
+    def backward(ctx, dout, *args):
+        x, weight, bias, seq_idx, initial_states = ctx.saved_tensors
+        dfinal_states = args[0] if ctx.return_final_states else None
+        if dout.stride(2) != 1 and dout.stride(1) != 1:
+            dout = dout.contiguous()
+        # The kernel supports passing in a pre-allocated dx (e.g., in case we want to fuse the
+        # backward of conv1d with the backward of chunk).
+        # Here we just pass in None and dx will be allocated in the C++ code.
+        dx, dweight, dbias, dinitial_states = causal_conv1d_bwd_function(
+            x,
+            weight,
+            bias,
+            dout,
+            seq_idx,
+            initial_states,
+            dfinal_states,
+            None,
+            ctx.return_dinitial_states,
+            ctx.activation,
+        )
+        return (
+            dx,
+            dweight,
+            dbias if bias is not None else None,
+            None,
+            dinitial_states if initial_states is not None else None,
+            None,
+            None,
+            None,
+        )
+
+
+def causal_conv1d_fn(
+    x,
+    weight,
+    bias=None,
+    seq_idx=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    seq_idx: (batch, seqlen)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1), to be written to
+    activation: either None or "silu" or "swish"
+
+    out: (batch, dim, seqlen)
+    """
+    return CausalConv1dFn.apply(
+        x,
+        weight,
+        bias,
+        seq_idx,
+        initial_states,
+        return_final_states,
+        final_states_out,
+        activation,
+    )
+
+
+def causal_conv1d_ref(
+    x,
+    weight,
+    bias=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1)
+
+    out: (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    x = x.to(weight.dtype)
+    seqlen = x.shape[-1]
+    dim, width = weight.shape
+    if initial_states is None:
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=width - 1, groups=dim)
+    else:
+        x = torch.cat([initial_states, x], dim=-1)
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=0, groups=dim)
+    out = out[..., :seqlen]
+    if return_final_states:
+        final_states = F.pad(x, (width - 1 - x.shape[-1], 0)).to(
+            dtype_in
+        )  # (batch, dim, width - 1)
+        if final_states_out is not None:
+            final_states_out.copy_(final_states)
+        else:
+            final_states_out = final_states
+    out = (out if activation is None else F.silu(out)).to(dtype=dtype_in)
+    return out if not return_final_states else (out, final_states_out)
+
+
+def causal_conv1d_update(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None, conv_state_indices=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len.
+    conv_state_indices: (batch,), dtype int32
+        If None, the conv_state is a larger tensor along the batch dim, 
+        and we are selecting the batch coords specified by conv_state_indices.
+        Useful for a continuous batching scenario.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    activation = activation in ["silu", "swish"]
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    out = causal_conv1d_update_function(
+        x, conv_state, weight, bias, activation, cache_seqlens, conv_state_indices
+    )
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return out
+
+
+def causal_conv1d_update_ref(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len before performing the convolution.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    batch, dim, seqlen = x.shape
+    width = weight.shape[1]
+    state_len = conv_state.shape[-1]
+    assert conv_state.shape == (batch, dim, state_len)
+    assert weight.shape == (dim, width)
+    if cache_seqlens is None:
+        x_new = torch.cat([conv_state, x], dim=-1).to(weight.dtype)  # (batch, dim, state_len + seqlen)
+        conv_state.copy_(x_new[:, :, -state_len:])
+    else:
+        width_idx = torch.arange(-(width - 1), 0, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        width_idx = torch.remainder(width_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        x_new = torch.cat([conv_state.gather(2, width_idx), x], dim=-1).to(weight.dtype)
+        copy_idx = torch.arange(seqlen, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        copy_idx = torch.remainder(copy_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        conv_state.scatter_(2, copy_idx, x)
+    out = F.conv1d(x_new, weight.unsqueeze(1), bias, padding=0, groups=dim)[:, :, -seqlen:]
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return (out if activation is None else F.silu(out)).to(dtype=dtype_in)

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/causal_conv1d_varlen.py

@@ -0,0 +1,86 @@
+import torch
+from torch import Tensor
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _causal_conv1d_varlen_states(
+    X,
+    CU_SEQLENS,
+    STATES,
+    state_len,
+    dim,
+    stride_x_seqlen, stride_x_dim,
+    stride_states_batch, stride_states_seqlen, stride_states_dim,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr
+):
+    batch_idx = tl.program_id(2)
+    STATES += batch_idx * stride_states_batch
+    end_idx = tl.load(CU_SEQLENS + batch_idx + 1)
+    start_idx = tl.maximum(tl.load(CU_SEQLENS + batch_idx), end_idx - state_len)
+    rows = end_idx - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    cols = tl.program_id(0) * BLOCK_N + tl.arange(0, BLOCK_N)
+    x = tl.load(X + rows[:, None] * stride_x_seqlen + cols[None, :] * stride_x_dim,
+                mask=(rows[:, None] >= start_idx) & (cols[None, :] < dim),
+                other=0)
+    rows_states = state_len - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    tl.store(STATES + rows_states[:, None] * stride_states_seqlen + cols[None, :] * stride_states_dim,
+             x,
+             mask=(rows_states[:, None] >= 0) & (cols[None, :] < dim))
+
+
+def causal_conv1d_varlen_states(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.empty(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    BLOCK_M = min(triton.next_power_of_2(state_len), 16)
+    BLOCK_N = min(triton.next_power_of_2(dim), 256)
+    grid = (triton.cdiv(dim, BLOCK_N), triton.cdiv(state_len, BLOCK_M), batch)
+    with torch.cuda.device(x.device.index):
+        _causal_conv1d_varlen_states[grid](
+            x,
+            cu_seqlens,
+            states,
+            state_len,
+            dim,
+            x.stride(0), x.stride(1),
+            states.stride(0), states.stride(2), states.stride(1),
+            BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N
+        )
+    return states
+
+
+def causal_conv1d_varlen_states_ref(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.zeros(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    for i in range(batch):
+        end_idx = cu_seqlens[i + 1]
+        start_idx = torch.maximum(cu_seqlens[i], end_idx - state_len)
+        states[i, :, -(end_idx - start_idx):] = x[start_idx:end_idx].T
+    return states

build/torch29-cxx11-cu128-aarch64-linux/causal_conv1d/cpp_functions.py

@@ -0,0 +1,96 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+
+from ._ops import ops
+
+def causal_conv1d_fwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    final_states_out: torch.Tensor | None,
+    silu_activation: bool,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_fwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        out=out,
+        final_states_out=final_states_out,
+        silu_activation=silu_activation,
+    )
+    return out
+
+
+def causal_conv1d_bwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    dout: torch.Tensor,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    dfinal_states: torch.Tensor | None,
+    dx: torch.Tensor | None,
+    return_dinitial_states: torch.Tensor,
+    silu_activation: bool,
+) -> tuple[torch.Tensor | None]:
+    batch_size, dim = x.size()[:2]
+    width = weight.size(-1)
+
+    if dx is None:
+        dx = torch.empty_like(x)
+    dweight = torch.zeros_like(weight, dtype=torch.float32)
+    dbias = None
+    if bias is not None:
+        dbias = torch.zeros_like(bias, dtype=torch.float32)
+    dinitial_states = None
+    if return_dinitial_states:
+        dinitial_states = torch.empty(batch_size, width - 1, dim, device=x.device, dtype=x.dtype).transpose(1, 2)
+
+    ops.causal_conv1d_bwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        dout=dout,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        dfinal_states=dfinal_states,
+        dx=dx,
+        dweight=dweight,
+        dbias=dbias,
+        dinitial_states=dinitial_states,
+        silu_activation=silu_activation,
+    )
+
+    dweight = dweight.type_as(weight)
+    if dbias is not None:
+        dbias = dbias.type_as(bias)
+    return dx, dweight, dbias, dinitial_states
+
+
+def causal_conv1d_update_function(
+    x: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    silu_activation: bool,
+    cache_seqlens: torch.Tensor | None,
+    conv_state_indices: torch.Tensor | None,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_update(
+        x=x,
+        conv_state=conv_state,
+        weight=weight,
+        bias=bias,
+        out=out,
+        silu_activation=silu_activation,
+        cache_seqlens=cache_seqlens,
+        conv_state_indices=conv_state_indices,
+    )
+    return out

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/__init__.py

@@ -0,0 +1,4 @@
+from .causal_conv1d_interface import causal_conv1d_fn, causal_conv1d_update
+from .causal_conv1d_varlen import causal_conv1d_varlen_states
+
+__all__ = ["causal_conv1d_fn", "causal_conv1d_update", "causal_conv1d_varlen_states"]

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/_causal_conv1d_306ae84.abi3.so

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a8f7ee76fbd178668417f36c383073d2aa2a20514ff05641f015ee9efaf6b1e6
+size 58331720

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/_ops.py

@@ -0,0 +1,9 @@
+import torch
+from . import _causal_conv1d_306ae84
+ops = torch.ops._causal_conv1d_306ae84
+
+def add_op_namespace_prefix(op_name: str):
+    """
+    Prefix op by namespace.
+    """
+    return f"_causal_conv1d_306ae84::{op_name}"

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/causal_conv1d_interface.py

@@ -0,0 +1,242 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+import torch.nn.functional as F
+
+from .cpp_functions import causal_conv1d_fwd_function, causal_conv1d_bwd_function, causal_conv1d_update_function
+
+
+class CausalConv1dFn(torch.autograd.Function):
+    @staticmethod
+    def forward(
+        ctx,
+        x,
+        weight,
+        bias=None,
+        seq_idx=None,
+        initial_states=None,
+        return_final_states=False,
+        final_states_out=None,
+        activation=None,
+    ):
+        if activation not in [None, "silu", "swish"]:
+            raise NotImplementedError("activation must be None, silu, or swish")
+        if x.stride(2) != 1 and x.stride(1) != 1:
+            x = x.contiguous()
+        bias = bias.contiguous() if bias is not None else None
+        if seq_idx is not None:
+            assert (
+                initial_states is None
+            ), "initial_states must be None if seq_idx is not None"
+            assert (
+                not return_final_states
+            ), "If seq_idx is not None, we don't return final_states_out"
+        seq_idx = seq_idx.contiguous() if seq_idx is not None else None
+        if initial_states is not None and (
+            initial_states.stride(2) != 1 and initial_states.stride(1) != 1
+        ):
+            initial_states = initial_states.contiguous()
+        if return_final_states:
+            assert (
+                x.stride(1) == 1
+            ), "Only channel-last layout support returning final_states_out"
+            if final_states_out is not None:
+                assert (
+                    final_states_out.stride(2) == 1 or final_states_out.stride(1) == 1
+                )
+            else:
+                batch, dim, seqlen = x.shape
+                width = weight.shape[1]
+                final_states_out = torch.empty(
+                    batch, width - 1, dim, device=x.device, dtype=x.dtype
+                ).transpose(1, 2)
+        else:
+            final_states_out = None
+        ctx.activation = activation in ["silu", "swish"]
+        out = causal_conv1d_fwd_function(
+            x, weight, bias, seq_idx, initial_states, final_states_out, ctx.activation
+        )
+        ctx.save_for_backward(x, weight, bias, seq_idx, initial_states)
+        ctx.return_final_states = return_final_states
+        ctx.return_dinitial_states = (
+            initial_states is not None and initial_states.requires_grad
+        )
+        return out if not return_final_states else (out, final_states_out)
+
+    @staticmethod
+    def backward(ctx, dout, *args):
+        x, weight, bias, seq_idx, initial_states = ctx.saved_tensors
+        dfinal_states = args[0] if ctx.return_final_states else None
+        if dout.stride(2) != 1 and dout.stride(1) != 1:
+            dout = dout.contiguous()
+        # The kernel supports passing in a pre-allocated dx (e.g., in case we want to fuse the
+        # backward of conv1d with the backward of chunk).
+        # Here we just pass in None and dx will be allocated in the C++ code.
+        dx, dweight, dbias, dinitial_states = causal_conv1d_bwd_function(
+            x,
+            weight,
+            bias,
+            dout,
+            seq_idx,
+            initial_states,
+            dfinal_states,
+            None,
+            ctx.return_dinitial_states,
+            ctx.activation,
+        )
+        return (
+            dx,
+            dweight,
+            dbias if bias is not None else None,
+            None,
+            dinitial_states if initial_states is not None else None,
+            None,
+            None,
+            None,
+        )
+
+
+def causal_conv1d_fn(
+    x,
+    weight,
+    bias=None,
+    seq_idx=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    seq_idx: (batch, seqlen)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1), to be written to
+    activation: either None or "silu" or "swish"
+
+    out: (batch, dim, seqlen)
+    """
+    return CausalConv1dFn.apply(
+        x,
+        weight,
+        bias,
+        seq_idx,
+        initial_states,
+        return_final_states,
+        final_states_out,
+        activation,
+    )
+
+
+def causal_conv1d_ref(
+    x,
+    weight,
+    bias=None,
+    initial_states=None,
+    return_final_states=False,
+    final_states_out=None,
+    activation=None,
+):
+    """
+    x: (batch, dim, seqlen)
+    weight: (dim, width)
+    bias: (dim,)
+    initial_states: (batch, dim, width - 1)
+    final_states_out: (batch, dim, width - 1)
+
+    out: (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    x = x.to(weight.dtype)
+    seqlen = x.shape[-1]
+    dim, width = weight.shape
+    if initial_states is None:
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=width - 1, groups=dim)
+    else:
+        x = torch.cat([initial_states, x], dim=-1)
+        out = F.conv1d(x, weight.unsqueeze(1), bias, padding=0, groups=dim)
+    out = out[..., :seqlen]
+    if return_final_states:
+        final_states = F.pad(x, (width - 1 - x.shape[-1], 0)).to(
+            dtype_in
+        )  # (batch, dim, width - 1)
+        if final_states_out is not None:
+            final_states_out.copy_(final_states)
+        else:
+            final_states_out = final_states
+    out = (out if activation is None else F.silu(out)).to(dtype=dtype_in)
+    return out if not return_final_states else (out, final_states_out)
+
+
+def causal_conv1d_update(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None, conv_state_indices=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len.
+    conv_state_indices: (batch,), dtype int32
+        If None, the conv_state is a larger tensor along the batch dim, 
+        and we are selecting the batch coords specified by conv_state_indices.
+        Useful for a continuous batching scenario.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    activation = activation in ["silu", "swish"]
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    out = causal_conv1d_update_function(
+        x, conv_state, weight, bias, activation, cache_seqlens, conv_state_indices
+    )
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return out
+
+
+def causal_conv1d_update_ref(x, conv_state, weight, bias=None, activation=None, cache_seqlens=None):
+    """
+    x: (batch, dim) or (batch, dim, seqlen)
+    conv_state: (batch, dim, state_len), where state_len >= width - 1
+    weight: (dim, width)
+    bias: (dim,)
+    cache_seqlens: (batch,), dtype int32.
+        If not None, the conv_state is treated as a circular buffer.
+        The conv_state will be updated by copying x to the conv_state starting at the index
+        @cache_seqlens % state_len before performing the convolution.
+
+    out: (batch, dim) or (batch, dim, seqlen)
+    """
+    if activation not in [None, "silu", "swish"]:
+        raise NotImplementedError("activation must be None, silu, or swish")
+    dtype_in = x.dtype
+    unsqueeze = x.dim() == 2
+    if unsqueeze:
+        x = x.unsqueeze(-1)
+    batch, dim, seqlen = x.shape
+    width = weight.shape[1]
+    state_len = conv_state.shape[-1]
+    assert conv_state.shape == (batch, dim, state_len)
+    assert weight.shape == (dim, width)
+    if cache_seqlens is None:
+        x_new = torch.cat([conv_state, x], dim=-1).to(weight.dtype)  # (batch, dim, state_len + seqlen)
+        conv_state.copy_(x_new[:, :, -state_len:])
+    else:
+        width_idx = torch.arange(-(width - 1), 0, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        width_idx = torch.remainder(width_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        x_new = torch.cat([conv_state.gather(2, width_idx), x], dim=-1).to(weight.dtype)
+        copy_idx = torch.arange(seqlen, dtype=torch.long, device=x.device).unsqueeze(0) + cache_seqlens.unsqueeze(1)
+        copy_idx = torch.remainder(copy_idx, state_len).unsqueeze(1).expand(-1, dim, -1)
+        conv_state.scatter_(2, copy_idx, x)
+    out = F.conv1d(x_new, weight.unsqueeze(1), bias, padding=0, groups=dim)[:, :, -seqlen:]
+    if unsqueeze:
+        out = out.squeeze(-1)
+    return (out if activation is None else F.silu(out)).to(dtype=dtype_in)

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/causal_conv1d_varlen.py

@@ -0,0 +1,86 @@
+import torch
+from torch import Tensor
+
+import triton
+import triton.language as tl
+
+
+@triton.jit
+def _causal_conv1d_varlen_states(
+    X,
+    CU_SEQLENS,
+    STATES,
+    state_len,
+    dim,
+    stride_x_seqlen, stride_x_dim,
+    stride_states_batch, stride_states_seqlen, stride_states_dim,
+    BLOCK_M: tl.constexpr,
+    BLOCK_N: tl.constexpr
+):
+    batch_idx = tl.program_id(2)
+    STATES += batch_idx * stride_states_batch
+    end_idx = tl.load(CU_SEQLENS + batch_idx + 1)
+    start_idx = tl.maximum(tl.load(CU_SEQLENS + batch_idx), end_idx - state_len)
+    rows = end_idx - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    cols = tl.program_id(0) * BLOCK_N + tl.arange(0, BLOCK_N)
+    x = tl.load(X + rows[:, None] * stride_x_seqlen + cols[None, :] * stride_x_dim,
+                mask=(rows[:, None] >= start_idx) & (cols[None, :] < dim),
+                other=0)
+    rows_states = state_len - (tl.program_id(1) + 1) * BLOCK_M + tl.arange(0, BLOCK_M)
+    tl.store(STATES + rows_states[:, None] * stride_states_seqlen + cols[None, :] * stride_states_dim,
+             x,
+             mask=(rows_states[:, None] >= 0) & (cols[None, :] < dim))
+
+
+def causal_conv1d_varlen_states(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.empty(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    BLOCK_M = min(triton.next_power_of_2(state_len), 16)
+    BLOCK_N = min(triton.next_power_of_2(dim), 256)
+    grid = (triton.cdiv(dim, BLOCK_N), triton.cdiv(state_len, BLOCK_M), batch)
+    with torch.cuda.device(x.device.index):
+        _causal_conv1d_varlen_states[grid](
+            x,
+            cu_seqlens,
+            states,
+            state_len,
+            dim,
+            x.stride(0), x.stride(1),
+            states.stride(0), states.stride(2), states.stride(1),
+            BLOCK_M=BLOCK_M, BLOCK_N=BLOCK_N
+        )
+    return states
+
+
+def causal_conv1d_varlen_states_ref(x: Tensor, cu_seqlens: Tensor, state_len: int) -> Tensor:
+    """
+    Forward pass only, does not support backward pass.
+    Parameters:
+        x: (total_tokens, dim)
+        cu_seqlens: (batch + 1), must already be sorted. The cumulative sum of the sequence lengths, starting from 0.
+        state_len: int. For each cu_seqlens, how many elements from x should be copied to the state.
+            If some of those elements belong to a different sequence, the value of the states will be zero.
+    Return:
+        states: (batch, dim, state_len)
+    """
+    _, dim = x.shape
+    batch = cu_seqlens.shape[0] - 1
+    cu_seqlens = cu_seqlens.contiguous()
+    states = torch.zeros(batch, state_len, dim, dtype=x.dtype, device=x.device).transpose(1, 2)
+    for i in range(batch):
+        end_idx = cu_seqlens[i + 1]
+        start_idx = torch.maximum(cu_seqlens[i], end_idx - state_len)
+        states[i, :, -(end_idx - start_idx):] = x[start_idx:end_idx].T
+    return states

build/torch29-cxx11-cu130-aarch64-linux/causal_conv1d/cpp_functions.py

@@ -0,0 +1,96 @@
+# Copyright (c) 2024, Tri Dao.
+
+import torch
+
+from ._ops import ops
+
+def causal_conv1d_fwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    final_states_out: torch.Tensor | None,
+    silu_activation: bool,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_fwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        out=out,
+        final_states_out=final_states_out,
+        silu_activation=silu_activation,
+    )
+    return out
+
+
+def causal_conv1d_bwd_function(
+    x: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    dout: torch.Tensor,
+    seq_idx: torch.Tensor | None,
+    initial_states: torch.Tensor | None,
+    dfinal_states: torch.Tensor | None,
+    dx: torch.Tensor | None,
+    return_dinitial_states: torch.Tensor,
+    silu_activation: bool,
+) -> tuple[torch.Tensor | None]:
+    batch_size, dim = x.size()[:2]
+    width = weight.size(-1)
+
+    if dx is None:
+        dx = torch.empty_like(x)
+    dweight = torch.zeros_like(weight, dtype=torch.float32)
+    dbias = None
+    if bias is not None:
+        dbias = torch.zeros_like(bias, dtype=torch.float32)
+    dinitial_states = None
+    if return_dinitial_states:
+        dinitial_states = torch.empty(batch_size, width - 1, dim, device=x.device, dtype=x.dtype).transpose(1, 2)
+
+    ops.causal_conv1d_bwd(
+        x=x,
+        weight=weight,
+        bias=bias,
+        dout=dout,
+        seq_idx=seq_idx,
+        initial_states=initial_states,
+        dfinal_states=dfinal_states,
+        dx=dx,
+        dweight=dweight,
+        dbias=dbias,
+        dinitial_states=dinitial_states,
+        silu_activation=silu_activation,
+    )
+
+    dweight = dweight.type_as(weight)
+    if dbias is not None:
+        dbias = dbias.type_as(bias)
+    return dx, dweight, dbias, dinitial_states
+
+
+def causal_conv1d_update_function(
+    x: torch.Tensor,
+    conv_state: torch.Tensor,
+    weight: torch.Tensor,
+    bias: torch.Tensor | None,
+    silu_activation: bool,
+    cache_seqlens: torch.Tensor | None,
+    conv_state_indices: torch.Tensor | None,
+) -> torch.Tensor:
+    out = torch.empty_like(x)
+    ops.causal_conv1d_update(
+        x=x,
+        conv_state=conv_state,
+        weight=weight,
+        bias=bias,
+        out=out,
+        silu_activation=silu_activation,
+        cache_seqlens=cache_seqlens,
+        conv_state_indices=conv_state_indices,
+    )
+    return out