sglang 0.5.2rc2__py3-none-any.whl → 0.5.3__py3-none-any.whl

sglang/srt/layers/attention/fla/fused_sigmoid_gating_recurrent.py

@@ -0,0 +1,232 @@
+from typing import Optional
+
+import torch
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.utils import input_guard
+
+
+@triton.heuristics(
+    {
+        "USE_INITIAL_STATE": lambda args: args["h0_source"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+@triton.jit(do_not_specialize=["T"])
+def fused_sigmoid_gating_delta_rule_update_kernel(
+    A_log,
+    a,
+    dt_bias,
+    softplus_beta,
+    softplus_threshold,
+    q,
+    k,
+    v,
+    b,
+    o,
+    h0_source,
+    h0_indices,
+    cu_seqlens,
+    scale,
+    T,
+    B: tl.constexpr,
+    H: tl.constexpr,
+    HV: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BK: tl.constexpr,
+    BV: tl.constexpr,
+    USE_INITIAL_STATE: tl.constexpr,
+    USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+):
+    """
+    Fused kernel that combines sigmoid gating computation with recurrent delta rule update.
+    """
+    i_k, i_v, i_nh = tl.program_id(0), tl.program_id(1), tl.program_id(2)
+    i_n, i_hv = i_nh // HV, i_nh % HV
+    i_h = i_hv // (HV // H)
+
+    if IS_VARLEN:
+        bos, eos = (
+            tl.load(cu_seqlens + i_n).to(tl.int64),
+            tl.load(cu_seqlens + i_n + 1).to(tl.int64),
+        )
+        all = T
+        T = eos - bos
+    else:
+        bos, eos = i_n * T, i_n * T + T
+        all = B * T
+
+    o_k = i_k * BK + tl.arange(0, BK)
+    o_v = i_v * BV + tl.arange(0, BV)
+
+    p_q = q + (bos * H + i_h) * K + o_k
+    p_k = k + (bos * H + i_h) * K + o_k
+    p_v = v + (bos * HV + i_hv) * V + o_v
+    p_b = b + bos * HV + i_hv
+    p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v
+
+    # Gating computation pointers
+    p_A_log = A_log + i_hv
+    p_a = a + bos * HV + i_hv
+    p_dt_bias = dt_bias + i_hv
+
+    mask_k = o_k < K
+    mask_v = o_v < V
+    mask_h = mask_k[:, None] & mask_v[None, :]
+
+    b_h = tl.zeros([BK, BV], dtype=tl.float32)
+    if USE_INITIAL_STATE:
+        idx = tl.load(h0_indices + i_n)
+        if idx >= 0:
+            p_h0 = (
+                h0_source
+                + idx * HV * K * V
+                + i_hv * K * V
+                + o_k[:, None] * V
+                + o_v[None, :]
+            )
+            b_h += tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
+
+    for _ in range(0, T):
+        # Load inputs
+        b_q = tl.load(p_q, mask=mask_k, other=0).to(tl.float32)
+        b_k = tl.load(p_k, mask=mask_k, other=0).to(tl.float32)
+        b_v = tl.load(p_v, mask=mask_v, other=0).to(tl.float32)
+        b_b = tl.load(p_b).to(tl.float32)
+
+        # Compute sigmoid gating
+        # Load gating parameters
+        b_A_log = tl.load(p_A_log).to(tl.float32)
+        b_a = tl.load(p_a).to(tl.float32)
+        b_dt_bias = tl.load(p_dt_bias).to(tl.float32)
+
+        # Compute g = -exp(A_log) * softplus(a + dt_bias)
+        x = b_a + b_dt_bias
+        beta_x = softplus_beta * x
+        # Apply softplus with numerical stability
+        softplus_x = tl.where(
+            beta_x <= softplus_threshold,
+            (1.0 / softplus_beta) * tl.log(1.0 + tl.exp(beta_x)),
+            x,
+        )
+        b_g = -tl.exp(b_A_log) * softplus_x
+
+        # Compute beta = sigmoid(b)
+        b_beta = 1.0 / (1.0 + tl.exp(-b_b))
+
+        # Apply L2 normalization if enabled
+        if USE_QK_L2NORM_IN_KERNEL:
+            b_q = b_q / (tl.sqrt(tl.sum(b_q * b_q) + 1e-6))
+            b_k = b_k / (tl.sqrt(tl.sum(b_k * b_k) + 1e-6))
+
+        b_q = b_q * scale
+
+        # Apply gating to hidden state: h *= exp(g)
+        b_h *= tl.exp(b_g)
+
+        # Delta rule: v -= sum(h * k, dim=0)
+        b_v -= tl.sum(b_h * b_k[:, None], 0)
+
+        # Apply beta gating: v *= beta
+        b_v *= b_beta
+
+        # Update hidden state: h += k[:, None] * v[None, :]
+        b_h += b_k[:, None] * b_v[None, :]
+
+        # Compute output: o = sum(h * q, dim=0)
+        b_o = tl.sum(b_h * b_q[:, None], 0)
+        tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
+
+        # Update pointers for next timestep
+        p_q += H * K
+        p_k += H * K
+        p_o += HV * V
+        p_v += HV * V
+        p_b += HV
+        p_a += HV
+
+    # Store final state back to h0_source with bounds checking
+    if USE_INITIAL_STATE:
+        idx = tl.load(h0_indices + i_n)
+        if idx >= 0:
+            p_h0 = (
+                h0_source
+                + idx * HV * K * V
+                + i_hv * K * V
+                + o_k[:, None] * V
+                + o_v[None, :]
+            )
+            tl.store(p_h0, b_h.to(p_h0.dtype.element_ty), mask=mask_h)
+
+
+@input_guard
+def fused_sigmoid_gating_delta_rule_update(
+    A_log: torch.Tensor,
+    a: torch.Tensor,
+    dt_bias: torch.Tensor,
+    softplus_beta: float,
+    softplus_threshold: float,
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    b: torch.Tensor,
+    initial_state_source: torch.Tensor,
+    initial_state_indices: torch.Tensor,
+    scale: Optional[float] = None,
+    use_qk_l2norm_in_kernel: bool = False,
+    cu_seqlens: Optional[torch.Tensor] = None,
+):
+    """
+    Fused triton implementation of sigmoid gating delta rule update.
+    This function uses a single fused kernel that combines both sigmoid gating computation
+    and the recurrent delta rule update for better performance.
+    """
+    B, T, H, K, V = *k.shape, v.shape[-1]
+    HV = v.shape[2]
+    N = B if cu_seqlens is None else len(cu_seqlens) - 1
+    BK, BV = triton.next_power_of_2(K), min(triton.next_power_of_2(V), 8)
+    NK, NV = triton.cdiv(K, BK), triton.cdiv(V, BV)
+    assert NK == 1, "NK > 1 is not supported yet"
+    num_stages = 3
+    num_warps = 1
+
+    if scale is None:
+        scale = k.shape[-1] ** -0.5
+    else:
+        assert scale > 0, "scale must be positive"
+
+    o = q.new_empty(NK, *v.shape)
+    grid = (NK, NV, N * HV)
+
+    fused_sigmoid_gating_delta_rule_update_kernel[grid](
+        A_log=A_log,
+        a=a,
+        dt_bias=dt_bias,
+        softplus_beta=softplus_beta,
+        softplus_threshold=softplus_threshold,
+        q=q,
+        k=k,
+        v=v,
+        b=b,
+        o=o,
+        h0_source=initial_state_source,
+        h0_indices=initial_state_indices,
+        cu_seqlens=cu_seqlens,
+        scale=scale,
+        T=T,
+        B=B,
+        H=H,
+        HV=HV,
+        K=K,
+        V=V,
+        BK=BK,
+        BV=BV,
+        USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+    o = o.squeeze(0)
+    return o

sglang/srt/layers/attention/fla/index.py

@@ -0,0 +1,37 @@
+# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/utils/index.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+import torch
+import torch.nn.functional as F
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.utils import tensor_cache
+
+
+@tensor_cache
+def prepare_lens(cu_seqlens: torch.LongTensor) -> torch.LongTensor:
+    return cu_seqlens[1:] - cu_seqlens[:-1]
+
+
+@tensor_cache
+def prepare_chunk_indices(
+    cu_seqlens: torch.LongTensor, chunk_size: int
+) -> torch.LongTensor:
+    indices = torch.cat(
+        [
+            torch.arange(n)
+            for n in triton.cdiv(prepare_lens(cu_seqlens), chunk_size).tolist()
+        ]
+    )
+    return torch.stack([indices.eq(0).cumsum(0) - 1, indices], 1).to(cu_seqlens)
+
+
+@tensor_cache
+def prepare_chunk_offsets(
+    cu_seqlens: torch.LongTensor, chunk_size: int
+) -> torch.LongTensor:
+    return torch.cat(
+        [cu_seqlens.new_tensor([0]), triton.cdiv(prepare_lens(cu_seqlens), chunk_size)]
+    ).cumsum(-1)

sglang/srt/layers/attention/fla/l2norm.py

@@ -0,0 +1,150 @@
+# Adapt from https://github.com/fla-org/flash-linear-attention/blob/main/fla/modules/l2norm.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional
+
+import torch
+import torch.nn as nn
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.utils import input_guard
+
+BT_LIST = [8, 16, 32, 64, 128]
+
+
+# @triton.autotune(
+#     configs=[
+#         triton.Config({}, num_warps=num_warps) for num_warps in [1, 2, 4, 8, 16, 32]
+#     ],
+#     key=["D"],
+# )
+@triton.jit
+def l2norm_fwd_kernel1(
+    x,
+    y,
+    D,
+    BD: tl.constexpr,
+    eps,
+):
+    i_t = tl.program_id(0)
+    x += i_t * D
+    y += i_t * D
+    # Compute mean and variance
+    cols = tl.arange(0, BD)
+    mask = cols < D
+    b_x = tl.load(x + cols, mask=mask, other=0.0).to(tl.float32)
+    b_var = tl.sum(b_x * b_x, axis=0)
+    b_rstd = 1 / tl.sqrt(b_var + eps)
+    # tl.store(Rstd + i_t, rstd)
+    # Normalize and apply linear transformation
+    b_y = b_x * b_rstd
+    tl.store(y + cols, b_y, mask=mask)
+
+
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BT": BT}, num_warps=num_warps)
+#         for num_warps in [1, 2, 4, 8, 16]
+#         for BT in BT_LIST
+#     ],
+#     key=["D", "NB"],
+# )
+@triton.jit
+def l2norm_fwd_kernel(
+    x,
+    y,
+    eps,
+    NB: tl.constexpr,
+    T: tl.constexpr,
+    D: tl.constexpr,
+    BT: tl.constexpr,
+    BD: tl.constexpr,
+):
+    i_t = tl.program_id(0)
+    p_x = tl.make_block_ptr(x, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0))
+    b_x = tl.load(p_x, boundary_check=(0, 1)).to(tl.float32)
+    b_var = tl.sum(b_x * b_x, axis=1)
+    b_y = b_x / tl.sqrt(b_var + eps)[:, None]
+    p_y = tl.make_block_ptr(y, (T, D), (D, 1), (i_t * BT, 0), (BT, BD), (1, 0))
+    tl.store(p_y, b_y.to(p_y.dtype.element_ty), boundary_check=(0, 1))
+
+
+def l2norm_fwd(
+    x: torch.Tensor, eps: float = 1e-6, output_dtype: Optional[torch.dtype] = None
+):
+    x_shape_og = x.shape
+    x = x.view(-1, x.shape[-1])
+    # allocate output
+    if output_dtype is None:
+        y = torch.empty_like(x)
+    else:
+        y = torch.empty_like(x, dtype=output_dtype)
+    assert y.stride(-1) == 1
+    T, D = x.shape[0], x.shape[-1]
+    # rstd = torch.empty((T,), dtype=torch.float32, device=x.device)
+    # Less than 64KB per feature: enqueue fused kernel
+    MAX_FUSED_SIZE = 65536 // x.element_size()
+    BD = min(MAX_FUSED_SIZE, triton.next_power_of_2(D))
+    if D > BD:
+        raise RuntimeError("This layer doesn't support feature dim >= 64KB.")
+
+    if D <= 512:
+        NB = triton.cdiv(T, 2048)
+
+        def grid(meta):
+            return (triton.cdiv(T, meta["BT"]),)
+
+        l2norm_fwd_kernel[grid](
+            x,
+            y,
+            eps,
+            NB=NB,
+            T=T,
+            D=D,
+            BD=BD,
+            BT=16,
+            num_warps=8,
+            num_stages=3,
+        )
+    else:
+        l2norm_fwd_kernel1[(T,)](
+            x,
+            y,
+            eps=eps,
+            D=D,
+            BD=BD,
+            num_warps=8,
+            num_stages=3,
+        )
+
+    return y.view(x_shape_og)
+
+
+class L2NormFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    def forward(ctx, x, eps=1e-6, output_dtype=None):
+        return l2norm_fwd(x, eps, output_dtype)
+
+
+def l2norm(
+    x: torch.Tensor, eps: float = 1e-6, output_dtype: Optional[torch.dtype] = None
+) -> torch.Tensor:
+    return L2NormFunction.apply(x, eps, output_dtype)
+
+
+l2_norm = l2norm
+
+
+class L2Norm(nn.Module):
+
+    def __init__(self, eps: float = 1e-6, output_dtype: Optional[torch.dtype] = None):
+        super().__init__()
+        self.eps = eps
+        self.output_dtype = output_dtype
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return l2norm(x, self.eps, self.output_dtype)