sglang 0.5.2rc2__py3-none-any.whl → 0.5.3.post1__py3-none-any.whl

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

@@ -0,0 +1,242 @@
+# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/gated_delta_rule/chunk.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+import warnings
+from typing import Optional
+
+import torch
+from einops import rearrange
+
+from sglang.srt.layers.attention.fla.chunk_delta_h import chunk_gated_delta_rule_fwd_h
+from sglang.srt.layers.attention.fla.chunk_o import chunk_fwd_o
+from sglang.srt.layers.attention.fla.chunk_scaled_dot_kkt import (
+    chunk_scaled_dot_kkt_fwd,
+)
+from sglang.srt.layers.attention.fla.cumsum import chunk_local_cumsum
+from sglang.srt.layers.attention.fla.l2norm import l2norm_fwd
+from sglang.srt.layers.attention.fla.solve_tril import solve_tril
+from sglang.srt.layers.attention.fla.utils import (
+    SUPPRESS_LEVEL,
+    autocast_custom_fwd,
+    input_guard,
+)
+from sglang.srt.layers.attention.fla.wy_fast import recompute_w_u_fwd
+
+
+def chunk_gated_delta_rule_fwd(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    g: torch.Tensor,
+    beta: torch.Tensor,
+    scale: float,
+    initial_state: torch.Tensor,
+    output_final_state: bool,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+):
+    g = chunk_local_cumsum(g, chunk_size=64, cu_seqlens=cu_seqlens)
+    # obtain WY representation. u is actually the new v.
+    A = chunk_scaled_dot_kkt_fwd(
+        k=k, beta=beta, g_cumsum=g, cu_seqlens=cu_seqlens, output_dtype=torch.float32
+    )
+    A = solve_tril(A=A, cu_seqlens=cu_seqlens, output_dtype=k.dtype)
+    w, u = recompute_w_u_fwd(
+        k=k,
+        v=v,
+        beta=beta,
+        A=A,
+        g_cumsum=g,
+        cu_seqlens=cu_seqlens,
+    )
+    h, v_new, final_state = chunk_gated_delta_rule_fwd_h(
+        k=k,
+        w=w,
+        u=u,
+        g=g,
+        initial_state=initial_state,
+        output_final_state=output_final_state,
+        cu_seqlens=cu_seqlens,
+    )
+    o = chunk_fwd_o(
+        q=q,
+        k=k,
+        v=v_new,
+        h=h,
+        g=g,
+        scale=scale,
+        cu_seqlens=cu_seqlens,
+    )
+    if SUPPRESS_LEVEL < 3:
+        return g, o, A, final_state, None, None, None
+    elif SUPPRESS_LEVEL >= 3:
+        return g, o, A, final_state, w, h, v_new
+
+
+class ChunkGatedDeltaRuleFunction(torch.autograd.Function):
+
+    @staticmethod
+    @input_guard
+    @autocast_custom_fwd
+    def forward(
+        ctx,
+        q: torch.Tensor,
+        k: torch.Tensor,
+        v: torch.Tensor,
+        g: torch.Tensor,
+        beta: torch.Tensor,
+        scale: float,
+        initial_state: torch.Tensor,
+        output_final_state: bool,
+        cu_seqlens: Optional[torch.LongTensor] = None,
+        use_qk_l2norm_in_kernel: bool = False,
+    ):
+        q_orig = q
+        k_orig = k
+
+        if use_qk_l2norm_in_kernel:
+            q = l2norm_fwd(q)
+            k = l2norm_fwd(k)
+
+        g, o, A, final_state, w, h, v_new = chunk_gated_delta_rule_fwd(
+            q=q,
+            k=k,
+            v=v,
+            g=g,
+            beta=beta,
+            scale=scale,
+            initial_state=initial_state,
+            output_final_state=output_final_state,
+            cu_seqlens=cu_seqlens,
+        )
+        return o.to(q.dtype), final_state
+
+
+@torch.compiler.disable
+def chunk_gated_delta_rule(
+    q: torch.Tensor,
+    k: torch.Tensor,
+    v: torch.Tensor,
+    g: torch.Tensor,
+    beta: torch.Tensor,
+    scale: float = None,
+    initial_state: torch.Tensor = None,
+    output_final_state: bool = False,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+    head_first: bool = False,
+    use_qk_l2norm_in_kernel: bool = False,
+):
+    r"""
+    Args:
+        q (torch.Tensor):
+            queries of shape `[B, T, H, K]` if `head_first=False` else `[B, H, T, K]`.
+        k (torch.Tensor):
+            keys of shape `[B, T, H, K]` if `head_first=False` else `[B, H, T, K]`.
+        v (torch.Tensor):
+            values of shape `[B, T, H, V]` if `head_first=False` else `[B, H, T, V]`.
+        g (torch.Tensor):
+            (forget) gating tensor (in log space!) of shape `[B, T, H]` if `head_first=False` else `[B, H, T]`.
+        beta (torch.Tensor):
+            betas of shape `[B, T, H]` if `head_first=False` else `[B, H, T]`.
+        scale (Optional[int]):
+            Scale factor for the RetNet attention scores.
+            If not provided, it will default to `1 / sqrt(K)`. Default: `None`.
+        initial_state (Optional[torch.Tensor]):
+            Initial state of shape `[N, H, K, V]` for `N` input sequences.
+            For equal-length input sequences, `N` equals the batch size `B`.
+            Default: `None`.
+        output_final_state (Optional[bool]):
+            Whether to output the final state of shape `[N, H, K, V]`. Default: `False`.
+        cu_seqlens (torch.LongTensor):
+            Cumulative sequence lengths of shape `[N+1]` used for variable-length training,
+            consistent with the FlashAttention API.
+        head_first (Optional[bool]):
+            Whether the inputs are in the head-first format, which is not supported for variable-length inputs.
+            Default: `False`.
+
+    Returns:
+        o (torch.Tensor):
+            Outputs of shape `[B, T, H, V]` if `head_first=False` else `[B, H, T, V]`.
+        final_state (torch.Tensor):
+            Final state of shape `[N, H, K, V]` if `output_final_state=True` else `None`.
+
+    Examples::
+        >>> import torch
+        >>> import torch.nn.functional as F
+        >>> from einops import rearrange
+        >>> from fla.ops.gated_delta_rule import chunk_gated_delta_rule
+        # inputs with equal lengths
+        >>> B, T, H, K, V = 4, 2048, 4, 512, 512
+        >>> q = torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda')
+        >>> k = F.normalize(torch.randn(B, T, H, K, dtype=torch.bfloat16, device='cuda'), p=2, dim=-1)
+        >>> v = torch.randn(B, T, H, V, dtype=torch.bfloat16, device='cuda')
+        >>> beta = torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda').sigmoid()
+        >>> g = F.logsigmoid(torch.rand(B, T, H, dtype=torch.bfloat16, device='cuda'))
+        >>> h0 = torch.randn(B, H, K, V, dtype=torch.bfloat16, device='cuda')
+        >>> o, ht = chunk_gated_delta_rule(
+            q, k, v, g, beta,
+            initial_state=h0,
+            output_final_state=True
+        )
+        # for variable-length inputs, the batch size `B` is expected to be 1 and `cu_seqlens` is required
+        >>> q, k, v, beta, g = map(lambda x: rearrange(x, 'b t ... -> 1 (b t) ...'), (q, k, v, beta, g))
+        # for a batch with 4 sequences, `cu_seqlens` with 5 start/end positions are expected
+        >>> cu_seqlens = q.new_tensor([0, 2048, 4096, 6144, 8192], dtype=torch.long)
+        >>> o_var, ht_var = chunk_gated_delta_rule(
+            q, k, v, g, beta,
+            initial_state=h0,
+            output_final_state=True,
+            cu_seqlens=cu_seqlens
+        )
+    """
+    assert q.dtype == k.dtype == v.dtype
+    assert (
+        q.dtype != torch.float32
+    ), "ChunkGatedDeltaRuleFunction does not support float32. Please use bfloat16."
+    assert (
+        len(beta.shape) == 3
+    ), "beta must be of shape [B, T, H] if head_first=False, or [B, H, T] otherwise."
+
+    if head_first:
+        raise DeprecationWarning(
+            "head_first is deprecated and will be removed in a future version. "
+            "Please use head_first=False for now instead."
+        )
+        q, k, v, beta, g = map(
+            lambda x: rearrange(x, "b h t ... -> b t h ..."), (q, k, v, beta, g)
+        )
+    # if not head_first and q.shape[1] < q.shape[2]:
+    #     warnings.warn(
+    #         f"Input tensor shape suggests potential format mismatch: seq_len ({q.shape[1]}) < num_heads ({q.shape[2]}). "
+    #         "This may indicate the inputs were passed in head-first format [B, H, T, ...] "
+    #         "when head_first=False was specified. "
+    #         "Please verify your input tensor format matches the expected shape [B, T, H, ...]."
+    #     )
+    if cu_seqlens is not None:
+        if q.shape[0] != 1:
+            raise ValueError(
+                f"The batch size is expected to be 1 rather than {q.shape[0]} when using `cu_seqlens`."
+                f"Please flatten variable-length inputs before processing."
+            )
+        if initial_state is not None and initial_state.shape[0] != len(cu_seqlens) - 1:
+            raise ValueError(
+                f"The number of initial states is expected to be equal to the number of input sequences, "
+                f"i.e., {len(cu_seqlens) - 1} rather than {initial_state.shape[0]}."
+            )
+    if scale is None:
+        scale = k.shape[-1] ** -0.5
+    o, final_state = ChunkGatedDeltaRuleFunction.apply(
+        q,
+        k,
+        v,
+        g,
+        beta,
+        scale,
+        initial_state,
+        output_final_state,
+        cu_seqlens,
+        use_qk_l2norm_in_kernel,
+    )
+    if head_first:
+        o = rearrange(o, "b t h ... -> b h t ...")
+    return o, final_state

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

@@ -0,0 +1,314 @@
+# Adapted from https://github.com/fla-org/flash-linear-attention/blob/main/fla/ops/common/chunk_delta_h.py
+# -*- coding: utf-8 -*-
+# Copyright (c) 2023-2025, Songlin Yang, Yu Zhang
+
+from typing import Optional, Tuple
+
+import torch
+import triton
+import triton.language as tl
+
+from sglang.srt.layers.attention.fla.index import (
+    prepare_chunk_indices,
+    prepare_chunk_offsets,
+)
+from sglang.srt.layers.attention.fla.op import exp, safe_exp
+from sglang.srt.layers.attention.fla.utils import is_nvidia_hopper
+
+NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8, 16]
+
+
+@triton.heuristics(
+    {
+        "USE_G": lambda args: args["g"] is not None,
+        "USE_INITIAL_STATE": lambda args: args["h0"] is not None,
+        "STORE_FINAL_STATE": lambda args: args["ht"] is not None,
+        "SAVE_NEW_VALUE": lambda args: args["v_new"] is not None,
+        "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
+    }
+)
+# @triton.autotune(
+#     configs=[
+#         triton.Config({"BV": BV}, num_warps=num_warps, num_stages=num_stages)
+#         for num_warps in [2, 4]
+#         for num_stages in [2, 3, 4]
+#         for BV in [32, 64]
+#     ],
+#     key=["H", "K", "V", "BT", "USE_G"],
+#     use_cuda_graph=use_cuda_graph,
+# )
+@triton.jit(do_not_specialize=["T"])
+def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
+    k,
+    v,
+    w,
+    v_new,
+    g,
+    h,
+    h0,
+    ht,
+    cu_seqlens,
+    chunk_offsets,
+    T,
+    H: tl.constexpr,
+    Hg: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BT: tl.constexpr,
+    BV: tl.constexpr,
+    USE_G: tl.constexpr,
+    USE_INITIAL_STATE: tl.constexpr,
+    STORE_FINAL_STATE: tl.constexpr,
+    SAVE_NEW_VALUE: tl.constexpr,
+    IS_VARLEN: tl.constexpr,
+):
+    i_v, i_nh = tl.program_id(0), tl.program_id(1)
+    i_n, i_h = i_nh // H, i_nh % H
+    if IS_VARLEN:
+        bos, eos = tl.load(cu_seqlens + i_n).to(tl.int32), tl.load(
+            cu_seqlens + i_n + 1
+        ).to(tl.int32)
+        T = eos - bos
+        NT = tl.cdiv(T, BT)
+        boh = tl.load(chunk_offsets + i_n).to(tl.int32)
+    else:
+        bos, eos = i_n * T, i_n * T + T
+        NT = tl.cdiv(T, BT)
+        boh = i_n * NT
+
+    # [BK, BV]
+    b_h1 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 64:
+        b_h2 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 128:
+        b_h3 = tl.zeros([64, BV], dtype=tl.float32)
+    if K > 192:
+        b_h4 = tl.zeros([64, BV], dtype=tl.float32)
+
+    # calculate offset
+    h += (boh * H + i_h) * K * V
+    v += (bos * H + i_h) * V
+    k += (bos * Hg + i_h // (H // Hg)) * K
+    w += (bos * H + i_h) * K
+    if SAVE_NEW_VALUE:
+        v_new += (bos * H + i_h) * V
+    stride_v = H * V
+    stride_h = H * K * V
+    stride_k = Hg * K
+    stride_w = H * K
+    if USE_INITIAL_STATE:
+        h0 = h0 + i_nh * K * V
+    if STORE_FINAL_STATE:
+        ht = ht + i_nh * K * V
+
+    # load initial state
+    if USE_INITIAL_STATE:
+        p_h0_1 = tl.make_block_ptr(h0, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0))
+        b_h1 += tl.load(p_h0_1, boundary_check=(0, 1)).to(tl.float32)
+        if K > 64:
+            p_h0_2 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h2 += tl.load(p_h0_2, boundary_check=(0, 1)).to(tl.float32)
+        if K > 128:
+            p_h0_3 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h3 += tl.load(p_h0_3, boundary_check=(0, 1)).to(tl.float32)
+        if K > 192:
+            p_h0_4 = tl.make_block_ptr(
+                h0, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            b_h4 += tl.load(p_h0_4, boundary_check=(0, 1)).to(tl.float32)
+
+    # main recurrence
+    for i_t in range(NT):
+        p_h1 = tl.make_block_ptr(
+            h + i_t * stride_h, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0)
+        )
+        tl.store(p_h1, b_h1.to(p_h1.dtype.element_ty), boundary_check=(0, 1))
+        if K > 64:
+            p_h2 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h2, b_h2.to(p_h2.dtype.element_ty), boundary_check=(0, 1))
+        if K > 128:
+            p_h3 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h3, b_h3.to(p_h3.dtype.element_ty), boundary_check=(0, 1))
+        if K > 192:
+            p_h4 = tl.make_block_ptr(
+                h + i_t * stride_h, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_h4, b_h4.to(p_h4.dtype.element_ty), boundary_check=(0, 1))
+
+        p_v = tl.make_block_ptr(
+            v, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+        )
+        p_v_new = (
+            tl.make_block_ptr(
+                v_new, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+            )
+            if SAVE_NEW_VALUE
+            else None
+        )
+        b_v_new = tl.zeros([BT, BV], dtype=tl.float32)
+        p_w = tl.make_block_ptr(
+            w, (T, K), (stride_w, 1), (i_t * BT, 0), (BT, 64), (1, 0)
+        )
+        b_w = tl.load(p_w, boundary_check=(0, 1))
+        b_v_new += tl.dot(b_w, b_h1.to(b_w.dtype))
+        if K > 64:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 64), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h2.to(b_w.dtype))
+        if K > 128:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 128), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h3.to(b_w.dtype))
+        if K > 192:
+            p_w = tl.make_block_ptr(
+                w, (T, K), (stride_w, 1), (i_t * BT, 192), (BT, 64), (1, 0)
+            )
+            b_w = tl.load(p_w, boundary_check=(0, 1))
+            b_v_new += tl.dot(b_w, b_h4.to(b_w.dtype))
+        b_v_new = -b_v_new + tl.load(p_v, boundary_check=(0, 1))
+
+        if SAVE_NEW_VALUE:
+            p_v_new = tl.make_block_ptr(
+                v_new, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+            )
+            tl.store(
+                p_v_new, b_v_new.to(p_v_new.dtype.element_ty), boundary_check=(0, 1)
+            )
+
+        if USE_G:
+            last_idx = min((i_t + 1) * BT, T) - 1
+            b_g_last = tl.load(g + bos * H + last_idx * H + i_h)
+            p_g = tl.make_block_ptr(
+                g + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)
+            )
+            b_g = tl.load(p_g, boundary_check=(0,))
+            b_v_new = b_v_new * safe_exp(b_g_last - b_g)[:, None]
+            b_g_last = exp(b_g_last)
+            b_h1 = b_h1 * b_g_last
+            if K > 64:
+                b_h2 = b_h2 * b_g_last
+            if K > 128:
+                b_h3 = b_h3 * b_g_last
+            if K > 192:
+                b_h4 = b_h4 * b_g_last
+        b_v_new = b_v_new.to(k.dtype.element_ty)
+        p_k = tl.make_block_ptr(
+            k, (K, T), (1, stride_k), (0, i_t * BT), (64, BT), (0, 1)
+        )
+        b_k = tl.load(p_k, boundary_check=(0, 1))
+        b_h1 += tl.dot(b_k, b_v_new)
+        if K > 64:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (64, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h2 += tl.dot(b_k, b_v_new)
+        if K > 128:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (128, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h3 += tl.dot(b_k, b_v_new)
+        if K > 192:
+            p_k = tl.make_block_ptr(
+                k, (K, T), (1, stride_k), (192, i_t * BT), (64, BT), (0, 1)
+            )
+            b_k = tl.load(p_k, boundary_check=(0, 1))
+            b_h4 += tl.dot(b_k, b_v_new)
+
+    # epilogue
+    if STORE_FINAL_STATE:
+        p_ht = tl.make_block_ptr(ht, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0))
+        tl.store(p_ht, b_h1.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 64:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (64, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h2.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 128:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (128, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h3.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+        if K > 192:
+            p_ht = tl.make_block_ptr(
+                ht, (K, V), (V, 1), (192, i_v * BV), (64, BV), (1, 0)
+            )
+            tl.store(p_ht, b_h4.to(p_ht.dtype.element_ty), boundary_check=(0, 1))
+
+
+def chunk_gated_delta_rule_fwd_h(
+    k: torch.Tensor,
+    w: torch.Tensor,
+    u: torch.Tensor,
+    g: Optional[torch.Tensor] = None,
+    initial_state: Optional[torch.Tensor] = None,
+    output_final_state: bool = False,
+    chunk_size: int = 64,  # SY: remove this argument and force chunk size 64?
+    save_new_value: bool = True,
+    cu_seqlens: Optional[torch.LongTensor] = None,
+) -> Tuple[torch.Tensor, torch.Tensor]:
+    B, T, Hg, K, V = *k.shape, u.shape[-1]
+    H = u.shape[-2]
+    BT = chunk_size
+
+    chunk_indices = (
+        prepare_chunk_indices(cu_seqlens, chunk_size)
+        if cu_seqlens is not None
+        else None
+    )
+    # N: the actual number of sequences in the batch with either equal or variable lengths
+    if cu_seqlens is None:
+        N, NT, chunk_offsets = B, triton.cdiv(T, BT), None
+    else:
+        N, NT, chunk_offsets = (
+            len(cu_seqlens) - 1,
+            len(chunk_indices),
+            prepare_chunk_offsets(cu_seqlens, BT),
+        )
+    assert K <= 256, "current kernel does not support head dimension larger than 256."
+
+    h = k.new_empty(B, NT, H, K, V)
+    final_state = (
+        k.new_empty(N, H, K, V, dtype=torch.float32) if output_final_state else None
+    )
+
+    v_new = torch.empty_like(u) if save_new_value else None
+
+    def grid(meta):
+        return (triton.cdiv(V, meta["BV"]), N * H)
+
+    chunk_gated_delta_rule_fwd_kernel_h_blockdim64[grid](
+        k=k,
+        v=u,
+        w=w,
+        v_new=v_new,
+        g=g,
+        h=h,
+        h0=initial_state,
+        ht=final_state,
+        cu_seqlens=cu_seqlens,
+        chunk_offsets=chunk_offsets,
+        T=T,
+        H=H,
+        Hg=Hg,
+        K=K,
+        V=V,
+        BT=BT,
+        BV=32,
+        num_warps=4,
+        num_stages=2,
+    )
+    return h, v_new, final_state