[FLA] Introduce Kimi Delta Attention(KDA) to VLLM

vllm/model_executor/layers/fla/ops/chunk.py

@@ -36,7 +36,7 @@ def chunk_gated_delta_rule_fwd(
     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
+        k=k, beta=beta, g=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(

vllm/model_executor/layers/fla/ops/chunk_delta_h.py

@@ -14,14 +14,15 @@
 
 from .index import prepare_chunk_indices, prepare_chunk_offsets
 from .op import exp
-from .utils import is_nvidia_hopper, use_cuda_graph
+from .utils import use_cuda_graph
 
-NUM_WARPS = [2, 4] if is_nvidia_hopper else [2, 4, 8, 16]
+NUM_WARPS = [2, 4, 8, 16]
 
 
 @triton.heuristics(
     {
         "USE_G": lambda args: args["g"] is not None,
+        "USE_GK": lambda args: args["gk"] 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,
@@ -35,7 +36,7 @@
         for num_stages in [2, 3, 4]
         for BV in [32, 64]
     ],
-    key=["H", "K", "V", "BT", "USE_G"],
+    key=["H", "K", "V", "BT"],
     use_cuda_graph=use_cuda_graph,
 )
 @triton.jit(do_not_specialize=["T"])
@@ -45,6 +46,7 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
     w,
     v_new,
     g,
+    gk,
     h,
     h0,
     ht,
@@ -58,6 +60,7 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
     BT: tl.constexpr,
     BV: tl.constexpr,
     USE_G: tl.constexpr,
+    USE_GK: tl.constexpr,
     USE_INITIAL_STATE: tl.constexpr,
     STORE_FINAL_STATE: tl.constexpr,
     SAVE_NEW_VALUE: tl.constexpr,
@@ -88,12 +91,12 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
         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
+    h += ((boh * H + i_h) * K * V).to(tl.int64)
+    v += ((bos * H + i_h) * V).to(tl.int64)
+    k += ((bos * Hg + i_h // (H // Hg)) * K).to(tl.int64)
+    w += ((bos * H + i_h) * K).to(tl.int64)
     if SAVE_NEW_VALUE:
-        v_new += (bos * H + i_h) * V
+        v_new += ((bos * H + i_h) * V).to(tl.int64)
     stride_v = H * V
     stride_h = H * K * V
     stride_k = Hg * K
@@ -145,92 +148,115 @@ def chunk_gated_delta_rule_fwd_kernel_h_blockdim64(
             )
             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))
+        b_v = 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))
+            b_v += 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))
+            b_v += 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))
+            b_v += tl.dot(b_w, b_h4.to(b_w.dtype))
+        p_v = tl.make_block_ptr(
+            v, (T, V), (stride_v, 1), (i_t * BT, i_v * BV), (BT, BV), (1, 0)
+        )
+        b_v = tl.load(p_v, boundary_check=(0, 1)) - b_v
 
         if SAVE_NEW_VALUE:
-            p_v_new = tl.make_block_ptr(
+            p_v = 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)
-            )
+            tl.store(p_v, b_v.to(p_v.dtype.element_ty), boundary_check=(0, 1))
 
+        last_idx = min((i_t + 1) * BT, T) - 1
         if USE_G:
             m_t = (i_t * BT + tl.arange(0, BT)) < T
-            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 * tl.where(m_t, exp(b_g_last - b_g), 0)[:, None]
+            b_v = b_v * tl.where(m_t, exp(b_g_last - b_g), 0)[:, None]
             b_g_last = exp(b_g_last)
-            b_h1 = b_h1 * b_g_last
+            b_h1 *= b_g_last
+            if K > 64:
+                b_h2 *= b_g_last
+            if K > 128:
+                b_h3 *= b_g_last
+            if K > 192:
+                b_h4 *= b_g_last
+
+        if USE_GK:
+            o_k1 = tl.arange(0, 64)
+            b_gk_last1 = tl.load(
+                gk + (bos + last_idx) * H * K + i_h * K + o_k1,
+                mask=(o_k1 < K),
+                other=0.0,
+            )
+            b_h1 *= exp(b_gk_last1)[:, None]
             if K > 64:
-                b_h2 = b_h2 * b_g_last
+                o_k2 = 64 + o_k1
+                b_gk_last2 = tl.load(
+                    gk + (bos + last_idx) * H * K + i_h * K + o_k2,
+                    mask=(o_k2 < K),
+                    other=0.0,
+                )
+                b_h2 *= exp(b_gk_last2)[:, None]
             if K > 128:
-                b_h3 = b_h3 * b_g_last
+                o_k3 = 128 + o_k1
+                b_gk_last3 = tl.load(
+                    gk + (bos + last_idx) * H * K + i_h * K + o_k3,
+                    mask=(o_k3 < K),
+                    other=0.0,
+                )
+                b_h3 *= exp(b_gk_last3)[:, None]
             if K > 192:
-                b_h4 = b_h4 * b_g_last
-        b_v_new = b_v_new.to(k.dtype.element_ty)
+                o_k4 = 192 + o_k1
+                b_gk_last4 = tl.load(
+                    gk + (bos + last_idx) * H * K + i_h * K + o_k4,
+                    mask=(o_k4 < K),
+                    other=0.0,
+                )
+                b_h4 *= exp(b_gk_last4)[:, None]
+        b_v = b_v.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)
+        b_h1 += tl.dot(b_k, b_v)
         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)
+            b_h2 += tl.dot(b_k, b_v)
         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)
+            b_h3 += tl.dot(b_k, b_v)
         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)
-
+            b_h4 += tl.dot(b_k, b_v)
     # epilogue
     if STORE_FINAL_STATE:
         p_ht = tl.make_block_ptr(ht, (K, V), (V, 1), (0, i_v * BV), (64, BV), (1, 0))
@@ -257,12 +283,15 @@ def chunk_gated_delta_rule_fwd_h(
     w: torch.Tensor,
     u: torch.Tensor,
     g: torch.Tensor | None = None,
+    gk: torch.Tensor | None = None,
     initial_state: torch.Tensor | None = 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: torch.LongTensor | None = None,
 ) -> tuple[torch.Tensor, torch.Tensor]:
+    # This kernel is slightly different from fla to support Q/K with different head numbers.
+    # In fla, Q/K always have the same head number, so Hg is always equal to H.
     B, T, Hg, K, V = *k.shape, u.shape[-1]
     H = u.shape[-2]
     BT = chunk_size
@@ -299,6 +328,7 @@ def grid(meta):
         w=w,
         v_new=v_new,
         g=g,
+        gk=gk,
         h=h,
         h0=initial_state,
         ht=final_state,

vllm/model_executor/layers/fla/ops/chunk_scaled_dot_kkt.py

@@ -18,8 +18,8 @@
 
 @triton.heuristics(
     {
+        "USE_G": lambda args: args["g"] is not None,
         "IS_VARLEN": lambda args: args["cu_seqlens"] is not None,
-        "USE_G": lambda args: args["g_cumsum"] is not None,
     }
 )
 @triton.autotune(
@@ -35,7 +35,7 @@
 def chunk_scaled_dot_kkt_fwd_kernel(
     k,
     beta,
-    g_cumsum,
+    g,
     A,
     cu_seqlens,
     chunk_indices,
@@ -85,9 +85,7 @@ def chunk_scaled_dot_kkt_fwd_kernel(
         b_A += tl.dot(b_kb.to(b_k.dtype), tl.trans(b_k))
 
     if USE_G:
-        p_g = tl.make_block_ptr(
-            g_cumsum + bos * H + i_h, (T,), (H,), (i_t * BT,), (BT,), (0,)
-        )
+        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_g_diff = b_g[:, None] - b_g[None, :]
         b_A = b_A * exp(b_g_diff)
@@ -102,8 +100,8 @@ def chunk_scaled_dot_kkt_fwd_kernel(
 
 def chunk_scaled_dot_kkt_fwd(
     k: torch.Tensor,
-    beta: torch.Tensor,
-    g_cumsum: torch.Tensor | None = None,
+    g: torch.Tensor | None = None,
+    beta: torch.Tensor | None = None,
     cu_seqlens: torch.LongTensor | None = None,
     chunk_size: int = 64,
     output_dtype: torch.dtype = torch.float32,
@@ -116,9 +114,8 @@ def chunk_scaled_dot_kkt_fwd(
             The key tensor of shape `[B, T, H, K]`.
         beta (torch.Tensor):
             The beta tensor of shape `[B, T, H]`.
-        g_cumsum (torch.Tensor):
-            The cumulative sum of the gate tensor of shape `[B, T, H]`.
-            Default: None
+        g (torch.Tensor):
+            The cumulative sum of the gate tensor of shape `[B, T, H]`. Default: `None`.
         cu_seqlens (torch.LongTensor):
             The cumulative sequence lengths of the input tensor.
             Default: None
@@ -130,20 +127,21 @@ def chunk_scaled_dot_kkt_fwd(
     Returns:
         beta * K * K^T of shape `[B, T, H, BT]` where `BT` is the chunk size.
     """
-
+    # This kernel is slightly different from fla to support Q/K with different head numbers.
+    # In fla, Q/K always have the same head number, so Hg is always equal to H.
     B, T, Hg, K = k.shape
-
     H = beta.shape[-1]
     BT = chunk_size
     chunk_indices = (
         prepare_chunk_indices(cu_seqlens, BT) if cu_seqlens is not None else None
     )
     NT = triton.cdiv(T, BT) if cu_seqlens is None else len(chunk_indices)
+
     A = torch.empty(B, T, H, BT, device=k.device, dtype=output_dtype)
     chunk_scaled_dot_kkt_fwd_kernel[(NT, B * H)](
         k=k,
+        g=g,
         beta=beta,
-        g_cumsum=g_cumsum,
         A=A,
         cu_seqlens=cu_seqlens,
         chunk_indices=chunk_indices,

vllm/model_executor/layers/fla/ops/fused_recurrent.py

@@ -57,6 +57,7 @@ def fused_recurrent_gated_delta_rule_fwd_kernel(
     IS_VARLEN: tl.constexpr,
     IS_CONTINUOUS_BATCHING: tl.constexpr,
     IS_SPEC_DECODING: tl.constexpr,
+    IS_KDA: tl.constexpr,
 ):
     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
@@ -86,7 +87,12 @@ def fused_recurrent_gated_delta_rule_fwd_kernel(
         p_beta = beta + (bos * HV + i_hv) * V + o_v
     else:
         p_beta = beta + bos * HV + i_hv
-    p_g = g + bos * HV + i_hv
+
+    if not IS_KDA:
+        p_g = g + bos * HV + i_hv
+    else:
+        p_gk = g + (bos * HV + i_hv) * K + o_k
+
     p_o = o + ((i_k * all + bos) * HV + i_hv) * V + o_v
 
     mask_k = o_k < K
@@ -116,14 +122,18 @@ def fused_recurrent_gated_delta_rule_fwd_kernel(
         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_g = tl.load(p_g).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
         # [BK, BV]
-        b_h *= exp(b_g)
+        if not IS_KDA:
+            b_g = tl.load(p_g).to(tl.float32)
+            b_h *= exp(b_g)
+        else:
+            b_gk = tl.load(p_gk).to(tl.float32)
+            b_h *= exp(b_gk[:, None])
         # [BV]
         b_v -= tl.sum(b_h * b_k[:, None], 0)
         if IS_BETA_HEADWISE:
@@ -155,7 +165,10 @@ def fused_recurrent_gated_delta_rule_fwd_kernel(
         p_k += H * K
         p_o += HV * V
         p_v += HV * V
-        p_g += HV
+        if not IS_KDA:
+            p_g += HV
+        else:
+            p_gk += HV * K
         p_beta += HV * (V if IS_BETA_HEADWISE else 1)
 
 
@@ -228,6 +241,7 @@ def fused_recurrent_gated_delta_rule_fwd(
         IS_BETA_HEADWISE=beta.ndim == v.ndim,
         USE_QK_L2NORM_IN_KERNEL=use_qk_l2norm_in_kernel,
         INPLACE_FINAL_STATE=inplace_final_state,
+        IS_KDA=False,
         num_warps=num_warps,
         num_stages=num_stages,
     )