[GDN] Support GDN packed decode (#20627)

benchmark/bench_linear_attention/bench_gdn_decode.py

@@ -0,0 +1,488 @@
+"""
+Benchmark & Correctness: GDN Packed Decode vs Baseline Decode.
+
+Compares:
+  - Baseline: split(mixed_qkv) → view → fused_sigmoid_gating_delta_rule_update
+  - Packed:   fused_recurrent_gated_delta_rule_packed_decode (single kernel)
+
+The packed path eliminates:
+  - torch.split() + .view() tensor materialization
+  - Separate gating kernel launches
+  - Intermediate tensor allocations
+
+Reports correctness (output & state matching) and performance (ms, speedup).
+
+Usage:
+    python bench_gdn_decode.py                        # default sweep
+    python bench_gdn_decode.py --mode bench           # benchmark only
+    python bench_gdn_decode.py --mode correctness     # correctness only
+    python bench_gdn_decode.py --preset qwen3.5-35b   # Qwen3.5-35B-A3B config
+"""
+
+import argparse
+import os
+import sys
+import time
+
+sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "python"))
+
+import torch
+import triton
+
+from sglang.srt.layers.attention.fla.fused_recurrent import (
+    fused_recurrent_gated_delta_rule_packed_decode,
+)
+from sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent import (
+    fused_sigmoid_gating_delta_rule_update,
+)
+
+# ---------------------------------------------------------------------------
+# Input factory
+# ---------------------------------------------------------------------------
+
+
+def make_inputs(
+    B: int,
+    H: int,
+    HV: int,
+    K: int,
+    V: int,
+    pool_size: int,
+    device: str,
+    dtype: torch.dtype,
+    seed: int = 42,
+):
+    """Create all input tensors for a single benchmark / correctness run."""
+    torch.manual_seed(seed)
+
+    qkv_dim = 2 * H * K + HV * V
+    mixed_qkv = torch.randn(B, qkv_dim, device=device, dtype=dtype)
+    a = torch.randn(B, HV, device=device, dtype=dtype)
+    b = torch.randn(B, HV, device=device, dtype=dtype)
+    A_log = torch.randn(HV, device=device, dtype=dtype)
+    dt_bias = torch.randn(HV, device=device, dtype=dtype)
+
+    ssm_states = torch.randn(pool_size, HV, V, K, device=device, dtype=dtype) * 0.1
+    cache_indices = torch.arange(B, device=device, dtype=torch.int32)
+
+    cu_seqlens = torch.arange(B + 1, device=device, dtype=torch.long)
+
+    return dict(
+        B=B,
+        H=H,
+        HV=HV,
+        K=K,
+        V=V,
+        qkv_dim=qkv_dim,
+        pool_size=pool_size,
+        mixed_qkv=mixed_qkv,
+        a=a,
+        b=b,
+        A_log=A_log,
+        dt_bias=dt_bias,
+        ssm_states=ssm_states,
+        cache_indices=cache_indices,
+        cu_seqlens=cu_seqlens,
+    )
+
+
+# ---------------------------------------------------------------------------
+# Runner wrappers
+# ---------------------------------------------------------------------------
+
+
+def run_baseline(inp):
+    """Baseline path: split → view → fused_sigmoid_gating_delta_rule_update.
+
+    This mirrors the FULL original decode path in GDNAttnBackend.forward_decode,
+    including the split, view, and kernel call.
+    """
+    B, H, HV, K, V = inp["B"], inp["H"], inp["HV"], inp["K"], inp["V"]
+    mixed_qkv = inp["mixed_qkv"]
+    ssm_states = inp["ssm_states"].clone()
+
+    # Step 1: split (same as forward_decode)
+    q_flat, k_flat, v_flat = torch.split(mixed_qkv, [H * K, H * K, HV * V], dim=-1)
+
+    # Step 2: view + reshape (same as forward_decode)
+    q = q_flat.view(1, B, H, K)
+    k = k_flat.view(1, B, H, K)
+    v = v_flat.view(1, B, HV, V)
+
+    # Step 3: fused gating + recurrent update
+    o = fused_sigmoid_gating_delta_rule_update(
+        A_log=inp["A_log"],
+        dt_bias=inp["dt_bias"],
+        q=q,
+        k=k,
+        v=v,
+        a=inp["a"],
+        b=inp["b"],
+        initial_state_source=ssm_states,
+        initial_state_indices=inp["cache_indices"],
+        cu_seqlens=inp["cu_seqlens"],
+        use_qk_l2norm_in_kernel=True,
+        softplus_beta=1.0,
+        softplus_threshold=20.0,
+    )
+
+    return o, ssm_states
+
+
+def run_packed(inp):
+    """Packed path: single fused kernel directly on mixed_qkv."""
+    B, HV, K, V = inp["B"], inp["HV"], inp["K"], inp["V"]
+    ssm_states = inp["ssm_states"].clone()
+    out = inp["mixed_qkv"].new_empty(B, 1, HV, V)
+
+    fused_recurrent_gated_delta_rule_packed_decode(
+        mixed_qkv=inp["mixed_qkv"],
+        a=inp["a"],
+        b=inp["b"],
+        A_log=inp["A_log"],
+        dt_bias=inp["dt_bias"],
+        scale=inp["K"] ** -0.5,
+        initial_state=ssm_states,
+        out=out,
+        ssm_state_indices=inp["cache_indices"],
+        use_qk_l2norm_in_kernel=True,
+    )
+
+    # Convert [B, 1, HV, V] → [1, B, HV, V] to match baseline layout
+    return out.transpose(0, 1), ssm_states
+
+
+# ---------------------------------------------------------------------------
+# Correctness check
+# ---------------------------------------------------------------------------
+
+
+def check_correctness(B, H, HV, K, V, pool_size, device, dtype, seed=42):
+    """Run correctness check for a single config. Returns True if PASS."""
+    tag = f"B={B:>4} H={H:>2} HV={HV:>2} K={K:>3} V={V:>3} pool={pool_size:>4}"
+
+    inp = make_inputs(B, H, HV, K, V, pool_size, device, dtype, seed=seed)
+
+    o_baseline, state_baseline = run_baseline(inp)
+    o_packed, state_packed = run_packed(inp)
+
+    # Output comparison
+    atol = 2e-2 if dtype != torch.float32 else 1e-4
+    rtol = 1e-2 if dtype != torch.float32 else 1e-4
+
+    try:
+        torch.testing.assert_close(o_packed, o_baseline, atol=atol, rtol=rtol)
+        output_ok = True
+    except AssertionError as e:
+        output_ok = False
+        out_diff = (o_packed - o_baseline).abs().max().item()
+
+    # State comparison (only for slots that were updated)
+    indices = inp["cache_indices"]
+    try:
+        torch.testing.assert_close(
+            state_packed[indices], state_baseline[indices], atol=atol, rtol=rtol
+        )
+        state_ok = True
+    except AssertionError:
+        state_ok = False
+        st_diff = (state_packed[indices] - state_baseline[indices]).abs().max().item()
+
+    passed = output_ok and state_ok
+
+    if passed:
+        print(f"  [PASS] {tag}")
+    else:
+        details = []
+        if not output_ok:
+            details.append(f"output max_diff={out_diff:.6f}")
+        if not state_ok:
+            details.append(f"state max_diff={st_diff:.6f}")
+        print(f"  [FAIL] {tag}  ({', '.join(details)})")
+
+    return passed
+
+
+# ---------------------------------------------------------------------------
+# Benchmark
+# ---------------------------------------------------------------------------
+
+
+def bench_shape(B, H, HV, K, V, pool_size, device, dtype):
+    """Benchmark baseline vs packed for a single config."""
+    inp = make_inputs(B, H, HV, K, V, pool_size, device, dtype)
+
+    # ── Baseline: full path including split + view ──
+    def fn_baseline():
+        q_flat, k_flat, v_flat = torch.split(
+            inp["mixed_qkv"], [H * K, H * K, HV * V], dim=-1
+        )
+        q = q_flat.view(1, B, H, K)
+        k = k_flat.view(1, B, H, K)
+        v = v_flat.view(1, B, HV, V)
+        fused_sigmoid_gating_delta_rule_update(
+            A_log=inp["A_log"],
+            dt_bias=inp["dt_bias"],
+            q=q,
+            k=k,
+            v=v,
+            a=inp["a"],
+            b=inp["b"],
+            initial_state_source=inp["ssm_states"],
+            initial_state_indices=inp["cache_indices"],
+            cu_seqlens=inp["cu_seqlens"],
+            use_qk_l2norm_in_kernel=True,
+            softplus_beta=1.0,
+            softplus_threshold=20.0,
+        )
+
+    # ── Packed: single kernel ──
+    out_buf = inp["mixed_qkv"].new_empty(B, 1, HV, V)
+
+    def fn_packed():
+        fused_recurrent_gated_delta_rule_packed_decode(
+            mixed_qkv=inp["mixed_qkv"],
+            a=inp["a"],
+            b=inp["b"],
+            A_log=inp["A_log"],
+            dt_bias=inp["dt_bias"],
+            scale=K**-0.5,
+            initial_state=inp["ssm_states"],
+            out=out_buf,
+            ssm_state_indices=inp["cache_indices"],
+            use_qk_l2norm_in_kernel=True,
+        )
+
+    # Warmup
+    for _ in range(10):
+        fn_baseline()
+        fn_packed()
+    torch.cuda.synchronize()
+
+    quantiles = [0.5, 0.2, 0.8]
+
+    try:
+        ms_baseline, ms_base_lo, ms_base_hi = triton.testing.do_bench(
+            fn_baseline, quantiles=quantiles, warmup=50, rep=200
+        )
+        ms_packed, ms_pack_lo, ms_pack_hi = triton.testing.do_bench(
+            fn_packed, quantiles=quantiles, warmup=50, rep=200
+        )
+    except Exception:
+        # Fallback to manual timing
+        torch.cuda.synchronize()
+        N = 200
+        start = time.perf_counter()
+        for _ in range(N):
+            fn_baseline()
+        torch.cuda.synchronize()
+        ms_baseline = (time.perf_counter() - start) / N * 1000
+
+        start = time.perf_counter()
+        for _ in range(N):
+            fn_packed()
+        torch.cuda.synchronize()
+        ms_packed = (time.perf_counter() - start) / N * 1000
+
+    speedup = ms_baseline / ms_packed if ms_packed > 0 else float("inf")
+    saved_us = (ms_baseline - ms_packed) * 1000
+
+    print(
+        f"  {B:>5}  {H:>3}  {HV:>3}  {K:>3}  {V:>3} | "
+        f"{ms_baseline * 1000:>10.1f} | "
+        f"{ms_packed * 1000:>10.1f} | "
+        f"{speedup:>7.2f}x | "
+        f"{saved_us:>+9.1f}"
+    )
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+
+def run_correctness(device, dtype):
+    print("=" * 70)
+    print("Correctness: Baseline GDN Decode vs Packed GDN Decode")
+    print("=" * 70)
+
+    shapes = [
+        # (B,   H,  HV,  K,   V,   pool_size)
+        # --- Qwen3.5-35B-A3B style (TP=2: H=8, HV=16) ---
+        (1, 8, 16, 128, 128, 32),
+        (4, 8, 16, 128, 128, 32),
+        (16, 8, 16, 128, 128, 64),
+        (32, 8, 16, 128, 128, 128),
+        (64, 8, 16, 128, 128, 128),
+        (128, 8, 16, 128, 128, 256),
+        (256, 8, 16, 128, 128, 512),
+        # --- Qwen3.5-35B-A3B style (TP=1: H=16, HV=32) ---
+        (1, 16, 32, 128, 128, 32),
+        (32, 16, 32, 128, 128, 128),
+        (64, 16, 32, 128, 128, 128),
+        # --- Qwen3-Next-80B-A3B style ---
+        (32, 16, 16, 128, 128, 128),
+        (64, 16, 16, 128, 128, 128),
+        # --- With PAD_SLOT_ID ---
+        (32, 8, 16, 128, 128, 128),  # some indices may be padded
+        # --- Edge cases ---
+        (1, 8, 16, 128, 128, 32),
+        (2, 8, 16, 128, 128, 32),
+    ]
+
+    all_pass = True
+    for B, H, HV, K, V, pool_size in shapes:
+        if not check_correctness(B, H, HV, K, V, pool_size, device, dtype):
+            all_pass = False
+
+    # PAD_SLOT_ID test
+    print("\n  PAD_SLOT_ID test (indices with -1):")
+    inp = make_inputs(32, 8, 16, 128, 128, 128, device, dtype)
+    o_baseline, st_baseline = run_baseline(inp)
+    o_packed, st_packed = run_packed(inp)
+
+    try:
+        torch.testing.assert_close(o_packed, o_baseline, atol=2e-2, rtol=1e-2)
+        print("  [PASS] PAD_SLOT_ID=-1 handling")
+    except AssertionError:
+        print("  [FAIL] PAD_SLOT_ID=-1 handling")
+        all_pass = False
+
+    print()
+    if all_pass:
+        print("ALL PASSED.")
+    else:
+        print("SOME FAILED.")
+    return all_pass
+
+
+def run_benchmark(device, dtype, args):
+    print()
+    print("=" * 85)
+    print("Benchmark: Baseline GDN Decode vs Packed GDN Decode")
+    print("=" * 85)
+
+    K = args.head_size_k
+    V = args.head_size_v
+    pool_size = args.pool_size
+
+    if args.preset == "qwen3.5-35b":
+        # Qwen3.5-35B-A3B: H_qk=16, H_v=32, K=128, V=128
+        # After TP=2: H=8, HV=16
+        bench_configs = [
+            # (B,   H,  HV) — TP=2 config
+            (1, 8, 16),
+            (2, 8, 16),
+            (4, 8, 16),
+            (8, 8, 16),
+            (16, 8, 16),
+            (32, 8, 16),
+            (64, 8, 16),
+            (128, 8, 16),
+            (256, 8, 16),
+            (512, 8, 16),
+            # TP=1 config (full heads)
+            (1, 16, 32),
+            (8, 16, 32),
+            (32, 16, 32),
+            (64, 16, 32),
+            (128, 16, 32),
+            (256, 16, 32),
+        ]
+    elif args.preset == "qwen3-next-80b":
+        bench_configs = [
+            # Qwen3-Next-80B-A3B: all same H=HV=16 after TP
+            (1, 16, 16),
+            (8, 16, 16),
+            (32, 16, 16),
+            (64, 16, 16),
+            (128, 16, 16),
+            (256, 16, 16),
+        ]
+    else:
+        bench_configs = []
+        for B in args.batch_sizes:
+            for H in args.num_q_heads:
+                for HV in args.num_v_heads:
+                    bench_configs.append((B, H, HV))
+
+    print(f"  Config: K={K}, V={V}, pool_size={pool_size}, dtype={dtype}")
+    print(
+        f"  {'B':>5}  {'H':>3}  {'HV':>3}  {'K':>3}  {'V':>3} | "
+        f"{'base (μs)':>10} | "
+        f"{'packed (μs)':>10} | "
+        f"{'speedup':>8} | "
+        f"{'saved (μs)':>10}"
+    )
+    print("  " + "-" * 75)
+
+    for B, H, HV in bench_configs:
+        actual_pool = max(pool_size, B + 16)
+        bench_shape(B, H, HV, K, V, actual_pool, device, dtype)
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description="Benchmark & Correctness: GDN Packed Decode vs Baseline"
+    )
+    parser.add_argument(
+        "--mode",
+        choices=["all", "correctness", "bench"],
+        default="all",
+        help="Run mode (default: all)",
+    )
+    parser.add_argument(
+        "--preset",
+        choices=["qwen3.5-35b", "qwen3-next-80b", "custom"],
+        default="qwen3.5-35b",
+        help="Preset config (default: qwen3.5-35b)",
+    )
+    parser.add_argument(
+        "--dtype",
+        choices=["float16", "bfloat16", "float32"],
+        default="bfloat16",
+    )
+    parser.add_argument("--head-size-k", type=int, default=128)
+    parser.add_argument("--head-size-v", type=int, default=128)
+    parser.add_argument("--pool-size", type=int, default=512)
+    parser.add_argument(
+        "--batch-sizes",
+        type=int,
+        nargs="+",
+        default=[1, 4, 8, 16, 32, 64, 128, 256, 512],
+    )
+    parser.add_argument(
+        "--num-q-heads",
+        type=int,
+        nargs="+",
+        default=[8, 16],
+    )
+    parser.add_argument(
+        "--num-v-heads",
+        type=int,
+        nargs="+",
+        default=[16, 32],
+    )
+    args = parser.parse_args()
+
+    device = "cuda"
+    dtype = getattr(torch, args.dtype)
+
+    cap = torch.cuda.get_device_capability()
+    dev_name = torch.cuda.get_device_name()
+    print(f"Device: {dev_name}  (SM {cap[0]}{cap[1]})")
+
+    if args.mode in ("all", "correctness"):
+        all_pass = run_correctness(device, dtype)
+        if not all_pass and args.mode == "all":
+            print("\nSkipping benchmark due to correctness failures.")
+            return 1
+
+    if args.mode in ("all", "bench"):
+        run_benchmark(device, dtype, args)
+
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

python/sglang/srt/layers/attention/fla/fused_recurrent.py

@@ -181,6 +181,227 @@ def fused_recurrent_gated_delta_rule_fwd(
     return o, final_state
 
 
+# Adapted from vllm project.
+@triton.jit
+def fused_recurrent_gated_delta_rule_packed_decode_kernel(
+    mixed_qkv,
+    a,
+    b,
+    A_log,
+    dt_bias,
+    o,
+    h0,
+    ht,
+    ssm_state_indices,
+    scale,
+    stride_mixed_qkv_tok: tl.constexpr,
+    stride_a_tok: tl.constexpr,
+    stride_b_tok: tl.constexpr,
+    stride_init_state_token: tl.constexpr,
+    stride_final_state_token: tl.constexpr,
+    stride_indices_seq: tl.constexpr,
+    H: tl.constexpr,
+    HV: tl.constexpr,
+    K: tl.constexpr,
+    V: tl.constexpr,
+    BK: tl.constexpr,
+    BV: tl.constexpr,
+    SOFTPLUS_THRESHOLD: tl.constexpr,
+    USE_QK_L2NORM_IN_KERNEL: tl.constexpr,
+):
+    i_v, i_nh = tl.program_id(0), tl.program_id(1)
+    i_n, i_hv = i_nh // HV, i_nh % HV
+    i_h = i_hv // (HV // H)
+
+    o_k = tl.arange(0, BK)
+    o_v = i_v * BV + tl.arange(0, BV)
+    mask_k = o_k < K
+    mask_v = o_v < V
+    mask_h = mask_v[:, None] & mask_k[None, :]
+
+    state_idx = tl.load(ssm_state_indices + i_n * stride_indices_seq).to(tl.int64)
+    p_o = o + (i_n * HV + i_hv) * V + o_v
+
+    if state_idx < 0:
+        zero = tl.zeros([BV], dtype=tl.float32).to(p_o.dtype.element_ty)
+        tl.store(p_o, zero, mask=mask_v)
+        return
+
+    p_h0 = h0 + state_idx * stride_init_state_token
+    p_h0 = p_h0 + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
+    b_h = tl.load(p_h0, mask=mask_h, other=0).to(tl.float32)
+
+    p_mixed = mixed_qkv + i_n * stride_mixed_qkv_tok
+    q_off = i_h * K + o_k
+    k_off = (H * K) + i_h * K + o_k
+    v_off = (2 * H * K) + i_hv * V + o_v
+    b_q = tl.load(p_mixed + q_off, mask=mask_k, other=0).to(tl.float32)
+    b_k = tl.load(p_mixed + k_off, mask=mask_k, other=0).to(tl.float32)
+    b_v = tl.load(p_mixed + v_off, mask=mask_v, other=0).to(tl.float32)
+
+    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
+
+    a_val = tl.load(a + i_n * stride_a_tok + i_hv).to(tl.float32)
+    b_val = tl.load(b + i_n * stride_b_tok + i_hv).to(tl.float32)
+    A_log_val = tl.load(A_log + i_hv).to(tl.float32)
+    dt_bias_val = tl.load(dt_bias + i_hv).to(tl.float32)
+    x = a_val + dt_bias_val
+    softplus_x = tl.where(x <= SOFTPLUS_THRESHOLD, tl.log(1.0 + tl.exp(x)), x)
+    g_val = -tl.exp(A_log_val) * softplus_x
+    beta_val = tl.sigmoid(b_val).to(b.dtype.element_ty).to(tl.float32)
+
+    b_h *= exp(g_val)
+    b_v -= tl.sum(b_h * b_k[None, :], 1)
+    b_v *= beta_val
+    b_h += b_v[:, None] * b_k[None, :]
+    b_o = tl.sum(b_h * b_q[None, :], 1)
+    tl.store(p_o, b_o.to(p_o.dtype.element_ty), mask=mask_v)
+
+    p_ht = ht + state_idx * stride_final_state_token
+    p_ht = p_ht + i_hv * V * K + o_v[:, None] * K + o_k[None, :]
+    tl.store(p_ht, b_h.to(p_ht.dtype.element_ty), mask=mask_h)
+
+
+def fused_recurrent_gated_delta_rule_packed_decode(
+    mixed_qkv: torch.Tensor,
+    a: torch.Tensor,
+    b: torch.Tensor,
+    A_log: torch.Tensor,
+    dt_bias: torch.Tensor,
+    scale: float,
+    initial_state: torch.Tensor,
+    out: torch.Tensor,
+    ssm_state_indices: torch.Tensor,
+    use_qk_l2norm_in_kernel: bool = False,
+) -> tuple[torch.Tensor, torch.Tensor]:
+    if mixed_qkv.ndim != 2:
+        raise ValueError(
+            f"`mixed_qkv` must be a 2D tensor (got ndim={mixed_qkv.ndim})."
+        )
+    if mixed_qkv.stride(-1) != 1:
+        raise ValueError("`mixed_qkv` must be contiguous in the last dim.")
+    if a.ndim != 2 or b.ndim != 2:
+        raise ValueError(
+            f"`a` and `b` must be 2D tensors (got a.ndim={a.ndim}, b.ndim={b.ndim})."
+        )
+    if a.stride(-1) != 1 or b.stride(-1) != 1:
+        raise ValueError("`a`/`b` must be contiguous in the last dim.")
+    if A_log.ndim != 1 or dt_bias.ndim != 1:
+        raise ValueError("`A_log`/`dt_bias` must be 1D tensors.")
+    if A_log.stride(0) != 1 or dt_bias.stride(0) != 1:
+        raise ValueError("`A_log`/`dt_bias` must be contiguous.")
+    if ssm_state_indices.ndim != 1:
+        raise ValueError(
+            f"`ssm_state_indices` must be 1D for packed decode (got ndim={ssm_state_indices.ndim})."
+        )
+    if not out.is_contiguous():
+        raise ValueError("`out` must be contiguous.")
+
+    dev = mixed_qkv.device
+    if any(
+        t.device != dev
+        for t in (a, b, A_log, dt_bias, initial_state, out, ssm_state_indices)
+    ):
+        raise ValueError("All inputs must be on the same device.")
+
+    B = mixed_qkv.shape[0]
+    if a.shape[0] != B or b.shape[0] != B:
+        raise ValueError(
+            "Mismatched batch sizes: "
+            f"mixed_qkv.shape[0]={B}, a.shape[0]={a.shape[0]}, b.shape[0]={b.shape[0]}."
+        )
+    if ssm_state_indices.shape[0] != B:
+        raise ValueError(
+            f"`ssm_state_indices` must have shape [B] (got {tuple(ssm_state_indices.shape)}; expected ({B},))."
+        )
+
+    if initial_state.ndim != 4:
+        raise ValueError(
+            f"`initial_state` must be a 4D tensor (got ndim={initial_state.ndim})."
+        )
+    if initial_state.stride(-1) != 1:
+        raise ValueError("`initial_state` must be contiguous in the last dim.")
+    HV, V, K = initial_state.shape[-3:]
+    if a.shape[1] != HV or b.shape[1] != HV:
+        raise ValueError(
+            f"`a`/`b` must have shape [B, HV] with HV={HV} (got a.shape={tuple(a.shape)}, b.shape={tuple(b.shape)})."
+        )
+    if A_log.numel() != HV or dt_bias.numel() != HV:
+        raise ValueError(
+            f"`A_log` and `dt_bias` must have {HV} elements (got A_log.numel()={A_log.numel()}, dt_bias.numel()={dt_bias.numel()})."
+        )
+    if out.shape != (B, 1, HV, V):
+        raise ValueError(
+            f"`out` must have shape {(B, 1, HV, V)} (got out.shape={tuple(out.shape)})."
+        )
+
+    qkv_dim = mixed_qkv.shape[1]
+    qk_dim = qkv_dim - HV * V
+    if qk_dim <= 0 or qk_dim % 2 != 0:
+        raise ValueError(
+            f"Invalid packed `mixed_qkv` last dim={qkv_dim} for HV={HV}, V={V}."
+        )
+    q_dim = qk_dim // 2
+    if q_dim % K != 0:
+        raise ValueError(f"Invalid packed Q size {q_dim}: must be divisible by K={K}.")
+    H = q_dim // K
+    if H <= 0 or HV % H != 0:
+        raise ValueError(
+            f"Invalid head config inferred from mixed_qkv: H={H}, HV={HV}."
+        )
+
+    BK = triton.next_power_of_2(K)
+    if triton.cdiv(K, BK) != 1:
+        raise ValueError(
+            f"Packed decode kernel only supports NK=1 (got K={K}, BK={BK})."
+        )
+    BV = min(triton.next_power_of_2(V), 32)
+    num_stages = 3
+    num_warps = 1
+
+    stride_mixed_qkv_tok = mixed_qkv.stride(0)
+    stride_a_tok = a.stride(0)
+    stride_b_tok = b.stride(0)
+    stride_init_state_token = initial_state.stride(0)
+    stride_final_state_token = initial_state.stride(0)
+    stride_indices_seq = ssm_state_indices.stride(0)
+
+    NV = triton.cdiv(V, BV)
+    grid = (NV, B * HV)
+    fused_recurrent_gated_delta_rule_packed_decode_kernel[grid](
+        mixed_qkv=mixed_qkv,
+        a=a,
+        b=b,
+        A_log=A_log,
+        dt_bias=dt_bias,
+        o=out,
+        h0=initial_state,
+        ht=initial_state,
+        ssm_state_indices=ssm_state_indices,
+        scale=scale,
+        stride_mixed_qkv_tok=stride_mixed_qkv_tok,
+        stride_a_tok=stride_a_tok,
+        stride_b_tok=stride_b_tok,
+        stride_init_state_token=stride_init_state_token,
+        stride_final_state_token=stride_final_state_token,
+        stride_indices_seq=stride_indices_seq,
+        H=H,
+        HV=HV,
+        K=K,
+        V=V,
+        BK=BK,
+        BV=BV,
+        SOFTPLUS_THRESHOLD=20.0,
+        USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
+        num_warps=num_warps,
+        num_stages=num_stages,
+    )
+    return out, initial_state
+
+
 class FusedRecurrentFunction(torch.autograd.Function):
 
     @staticmethod

python/sglang/srt/layers/attention/linear/gdn_backend.py

@@ -1,4 +1,4 @@
-from typing import Tuple, Union
+from typing import Optional, Tuple, Union
 
 import torch
 
@@ -111,10 +111,48 @@ class GDNKernelDispatcher:
         else:
             self.verify_kernel = triton_kernel
 
+        self.supports_packed_decode = getattr(
+            self.decode_kernel, "supports_packed_decode", False
+        )
+
         rank0_log(
             f"GDN kernel dispatcher: decode={self.decode_kernel.__class__.__name__}, "
             f"extend={self.extend_kernel.__class__.__name__}, "
             f"verify={self.verify_kernel.__class__.__name__} "
+            f"packed_decode={self.supports_packed_decode}"
+        )
+
+    def packed_decode(
+        self,
+        mixed_qkv: torch.Tensor,
+        a: torch.Tensor,
+        b: torch.Tensor,
+        *,
+        A_log: torch.Tensor,
+        dt_bias: torch.Tensor,
+        scale: float,
+        ssm_states: torch.Tensor,
+        cache_indices: torch.Tensor,
+        num_v_heads: int,
+        head_v_dim: int,
+        **kwargs,
+    ) -> Optional[torch.Tensor]:
+        """Attempt packed decode. Returns output tensor or None if
+        the decode kernel does not support packed decode."""
+        if not self.supports_packed_decode:
+            return None
+        return self.decode_kernel.packed_decode(
+            mixed_qkv,
+            a,
+            b,
+            A_log=A_log,
+            dt_bias=dt_bias,
+            scale=scale,
+            ssm_states=ssm_states,
+            cache_indices=cache_indices,
+            num_v_heads=num_v_heads,
+            head_v_dim=head_v_dim,
+            **kwargs,
         )
 
     def decode(
@@ -243,6 +281,26 @@ class GDNAttnBackend(MambaAttnBackendBase):
             conv_state_indices=cache_indices,
         )
 
+        # Skip split + reshape + separate gating kernel by consuming
+        # the packed mixed_qkv directly in a single fused Triton kernel.
+        if self.kernel_dispatcher.supports_packed_decode:
+            core_attn_out = self.kernel_dispatcher.packed_decode(
+                mixed_qkv=mixed_qkv,
+                a=a,
+                b=b,
+                A_log=layer.A_log,
+                dt_bias=layer.dt_bias,
+                scale=layer.head_k_dim**-0.5,
+                ssm_states=ssm_states,
+                cache_indices=cache_indices,
+                num_v_heads=layer.num_v_heads,
+                head_v_dim=layer.head_v_dim,
+            )
+            self._track_mamba_state_decode(
+                forward_batch, conv_states, ssm_states, cache_indices
+            )
+            return core_attn_out
+
         query, key, value = torch.split(
             mixed_qkv,
             [layer.q_dim, layer.k_dim, layer.v_dim],

python/sglang/srt/layers/attention/linear/kernels/gdn_triton.py

@@ -7,6 +7,9 @@ from sglang.srt.utils import is_cpu, is_npu
 
 if not is_cpu():
     from sglang.srt.layers.attention.fla.chunk import chunk_gated_delta_rule
+    from sglang.srt.layers.attention.fla.fused_recurrent import (
+        fused_recurrent_gated_delta_rule_packed_decode,
+    )
     from sglang.srt.layers.attention.fla.fused_sigmoid_gating_recurrent import (
         fused_sigmoid_gating_delta_rule_update,
     )
@@ -31,6 +34,63 @@ elif is_cpu():
 class TritonGDNKernel(LinearAttnKernelBase):
     """Triton-based kernel for GDN (Gated Delta Network) linear attention."""
 
+    supports_packed_decode: bool = not is_cpu() and not is_npu()
+
+    def packed_decode(
+        self,
+        mixed_qkv: torch.Tensor,
+        a: torch.Tensor,
+        b: torch.Tensor,
+        *,
+        A_log: torch.Tensor,
+        dt_bias: torch.Tensor,
+        scale: float,
+        ssm_states: torch.Tensor,
+        cache_indices: torch.Tensor,
+        num_v_heads: int,
+        head_v_dim: int,
+        **kwargs,
+    ) -> torch.Tensor:
+        """Packed decode fast path: fuse QKV extraction + gating + recurrent
+        update into a single Triton kernel, eliminating intermediate tensors
+        and extra kernel launches.
+
+        Args:
+            mixed_qkv: [B, qkv_dim] packed projection output after conv1d.
+            a, b: [B, HV] gating inputs.
+            A_log: [HV] log-space decay parameter.
+            dt_bias: [HV] time-step bias.
+            scale: attention scale factor (typically head_k_dim ** -0.5).
+            ssm_states: [num_slots, HV, V, K] full state pool.
+            cache_indices: [B] per-request state slot indices.
+            num_v_heads: number of value heads (after TP sharding).
+            head_v_dim: dimension per value head.
+
+        Returns:
+            output tensor of shape [1, B, HV, V] matching the existing
+            decode kernel output layout.
+        """
+        B = mixed_qkv.shape[0]
+        # Packed kernel expects output shape [B, 1, HV, V]
+        out = mixed_qkv.new_empty(B, 1, num_v_heads, head_v_dim)
+
+        fused_recurrent_gated_delta_rule_packed_decode(
+            mixed_qkv=mixed_qkv,
+            a=a,
+            b=b,
+            A_log=A_log,
+            dt_bias=dt_bias,
+            scale=scale,
+            initial_state=ssm_states,
+            out=out,
+            ssm_state_indices=cache_indices,
+            use_qk_l2norm_in_kernel=True,
+        )
+
+        # Convert [B, 1, HV, V] → [1, B, HV, V] to match existing output
+        # layout. transpose() returns a view — zero cost.
+        return out.transpose(0, 1)
+
     def decode(
         self,
         q: torch.Tensor,