@novastera-oss/llamarn 0.2.9 → 0.3.1

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/conv2d_mm.comp

@@ -0,0 +1,265 @@
+#version 450
+
+#ifdef USE_COLLECTIVES
+#    extension GL_KHR_shader_subgroup_shuffle : enable
+#endif
+
+#include "types.comp"
+
+// Make spec constant
+#define SHMEM_PAD 0
+
+// shape notation: [dim(N), ..., dim(0)] -- stride(dim(j)) >= stride(dim(i)) if i > j
+layout(binding = 0) readonly buffer A {
+    A_TYPE knl_data[];
+};  // src0 - kernel:   [KW, KH, Cin, Cout]
+
+layout(binding = 1) readonly buffer B {
+    B_TYPE src_data[];
+};  // src1 - input:    [W, H, Cin, N] -- channel_first format
+
+layout(binding = 2) writeonly buffer D {
+    D_TYPE dst_data[];
+};  // dst - result:    [OW, OH, Cout, N]
+
+layout(push_constant) uniform parameter {
+    // I/O channels, batch size
+    uint32_t Cout;
+    uint32_t Cin;
+    uint32_t N;
+
+    // Tensor spatial sizes: kernel, input, output
+    uint32_t KW;
+    uint32_t KH;
+    uint32_t W;
+    uint32_t H;
+    uint32_t OW;
+    uint32_t OH;
+
+    // Parameters: stride, padding, dilation - 0=y, 1=x
+    uint32_t s0;
+    uint32_t s1;
+    uint32_t p0;
+    uint32_t p1;
+    uint32_t d0;
+    uint32_t d1;
+
+    // Strides in elements
+    uint32_t nb01;
+    uint32_t nb02;
+    uint32_t nb03;
+
+    uint32_t nb11;
+    uint32_t nb12;
+    uint32_t nb13;
+
+    uint32_t nb1;
+    uint32_t nb2;
+    uint32_t nb3;
+}
+
+p;
+
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+// Blocktile sizes
+layout(constant_id = 1) const uint BS_K            = 128;
+layout(constant_id = 2) const uint BS_CRS          = 16;
+layout(constant_id = 3) const uint BS_NPQ          = 128;
+// Thread-tile sizes
+layout(constant_id = 4) const uint TS_K            = 8;
+layout(constant_id = 5) const uint use_collectives = 1;
+
+uint32_t       tid     = gl_LocalInvocationID.x;
+const uint32_t WG_SIZE = gl_WorkGroupSize.x;
+
+uint splitWork(uint work_size, uint block_size) {
+    return (block_size + work_size - 1) / block_size;
+}
+
+uint32_t K   = p.Cout;
+uint32_t CRS = p.Cin * p.KH * p.KW;
+uint32_t NPQ = p.N * p.OH * p.OW;
+
+uint32_t n_elems_out = K * NPQ;
+
+// Number of blocktiles per input
+uint32_t NB_CRS = splitWork(CRS, BS_CRS);
+
+const uint32_t Ash_stride = BS_CRS + SHMEM_PAD;
+const uint32_t Bsh_stride = BS_NPQ + SHMEM_PAD;
+
+const uint32_t Ash_numel = BS_K * BS_CRS;
+const uint32_t Bsh_numel = BS_CRS * BS_NPQ;
+
+const uint32_t Ash_len = BS_K * Ash_stride;
+const uint32_t Bsh_len = BS_CRS * Bsh_stride;
+
+shared float Ash[Ash_len];  // K x CRS
+shared float Bsh[Bsh_len];  // CRS x NPQ
+
+// Threadtile sizes
+const uint32_t TS_NPQ = BS_K * BS_NPQ / WG_SIZE / TS_K;
+
+// Number of threadtiles per blocktile
+const uint32_t NT_K   = BS_K / TS_K;
+const uint32_t NT_NPQ = BS_NPQ / TS_NPQ;
+
+float regA[TS_K];
+float regB[TS_NPQ];
+float regC[TS_K][TS_NPQ];
+
+/*
+Compute
+KxCRS @ CRSxNPQ = K x NPQ
+K=Cout
+C=Cin
+R,S=KH,KW
+P,Q=OH,OW
+*/
+
+uint32_t B_idx_K   = gl_WorkGroupID.x;
+uint32_t B_idx_NPQ = gl_WorkGroupID.y;
+
+uint32_t T_y = tid / NT_NPQ;
+uint32_t T_x = tid % NT_NPQ;
+
+uint32_t       Ar    = tid / BS_CRS;
+uint32_t       Ac    = tid % BS_CRS;
+const uint32_t ArpWg = WG_SIZE / BS_CRS;
+
+uint32_t       Br    = tid / BS_NPQ;
+uint32_t       Bc    = tid % BS_NPQ;
+const uint32_t BrpWg = WG_SIZE / BS_NPQ;
+
+void main() {
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            regC[T_ly][T_lx] = 0.0;
+        }
+    }
+    /* Advance block in CRS dim */
+    for (uint32_t B_idx_CRS = 0; B_idx_CRS < NB_CRS; B_idx_CRS++) {
+        uint32_t CRS_idx_a;
+        uint32_t Cin_idx_a;
+        uint32_t KH_idx_a;
+        uint32_t KW_idx_a;
+
+#ifdef USE_COLLECTIVES
+        uint32_t cached_CRS_idx;
+        uint32_t cached_Cin_idx;
+        uint32_t cached_KH_idx;
+        uint32_t cached_KW_idx;
+        if (use_collectives == 1) {
+            cached_CRS_idx                = B_idx_CRS * BS_CRS + gl_SubgroupInvocationID;
+            cached_Cin_idx                = cached_CRS_idx / (p.KW * p.KH);
+            uint32_t cached_CRS_remainder = (cached_CRS_idx - cached_Cin_idx * p.KW * p.KH);
+            cached_KH_idx                 = cached_CRS_remainder / p.KW;
+            cached_KW_idx                 = cached_CRS_remainder - cached_KH_idx * p.KW;
+
+            CRS_idx_a = subgroupShuffle(cached_CRS_idx, Ac);
+            Cin_idx_a = subgroupShuffle(cached_Cin_idx, Ac);
+            KH_idx_a  = subgroupShuffle(cached_KH_idx, Ac);
+            KW_idx_a  = subgroupShuffle(cached_KW_idx, Ac);
+        } else {
+            CRS_idx_a              = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+            Cin_idx_a              = CRS_idx_a / (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+            KH_idx_a               = CRS_remainder / p.KW;
+            KW_idx_a               = CRS_remainder - KH_idx_a * p.KW;
+        }
+#else
+        CRS_idx_a     = B_idx_CRS * BS_CRS + Ac;  // Global CRS_idx_a (column index of A)
+        Cin_idx_a     = CRS_idx_a / (p.KW * p.KH);
+        CRS_remainder = CRS_idx_a - Cin_idx_a * p.KW * p.KH;
+        KH_idx_a      = CRS_remainder / p.KW;
+        KW_idx_a      = CRS_remainder - KH_idx_a * p.KW;
+#endif
+
+        /* Load kernel to A_block: (BS_K x BS_CRS)*/
+        for (uint32_t r_offset = 0; r_offset < BS_K; r_offset += ArpWg) {
+            uint32_t B_ly    = r_offset + Ar;
+            uint32_t B_lx    = Ac;
+            uint32_t K_idx   = B_idx_K * BS_K + B_ly; /* Global K_idx (row index of A)*/
+            uint32_t knl_idx = min(KW_idx_a + KH_idx_a * p.nb01 + Cin_idx_a * p.nb02 + K_idx * p.nb03, K * CRS - 1);
+            float    val     = knl_data[knl_idx];
+            if (K_idx >= K || CRS_idx_a >= CRS) {
+                val = 0.0;
+            }
+            Ash[B_ly * Ash_stride + B_lx] = val;
+        }
+        /* Load input to B_block: (BS_CRS x BS_NPQ) */
+        for (uint32_t r_offset = 0; r_offset < BS_CRS; r_offset += BrpWg) {
+            uint32_t B_ly          = r_offset + Br;             /* Row index of B block */
+            uint32_t B_lx          = Bc;
+            uint32_t NPQ_idx       = B_idx_NPQ * BS_NPQ + B_lx; /* Global NPQ index (column index of B) */
+            uint32_t N_idx         = NPQ_idx / (p.OH * p.OW);
+            uint32_t NPQ_remainder = NPQ_idx - N_idx * p.OH * p.OW;
+            uint32_t OH_idx        = NPQ_remainder / p.OW;
+            uint32_t OW_idx        = NPQ_remainder - OH_idx * p.OW;
+
+            uint32_t CRS_idx_b;
+            uint32_t Cin_idx_b;
+            uint32_t KH_idx_b;
+            uint32_t KW_idx_b;
+#ifdef USE_COLLECTIVES
+            if (use_collectives == 1) {
+                CRS_idx_b = subgroupShuffle(cached_CRS_idx, r_offset + Br);
+                Cin_idx_b = subgroupShuffle(cached_Cin_idx, r_offset + Br);
+                KH_idx_b  = subgroupShuffle(cached_KH_idx, r_offset + Br);
+                KW_idx_b  = subgroupShuffle(cached_KW_idx, r_offset + Br);
+            } else {
+                CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+                Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+                uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+                KH_idx_b               = CRS_remainder / p.KW;
+                KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+            }
+#else
+            CRS_idx_b              = B_idx_CRS * BS_CRS + B_ly; /* Global CRS index (row index of B) */
+            Cin_idx_b              = CRS_idx_b / (p.KW * p.KH);
+            uint32_t CRS_remainder = CRS_idx_b - Cin_idx_b * p.KW * p.KH;
+            KH_idx_b               = CRS_remainder / p.KW;
+            KW_idx_b               = CRS_remainder - KH_idx_b * p.KW;
+#endif
+
+            uint32_t H_idx = OH_idx * p.s1 + KH_idx_b * p.d1 - p.p1;
+            uint32_t W_idx = OW_idx * p.s0 + KW_idx_b * p.d0 - p.p0;
+            uint32_t src_idx =
+                min(max(W_idx + H_idx * p.nb11 + Cin_idx_b * p.nb12 + N_idx * p.nb13, 0), p.Cin * p.N * p.W * p.H - 1);
+            float val = src_data[src_idx];
+            if (CRS_idx_b >= CRS || NPQ_idx >= NPQ || H_idx < 0 || H_idx >= p.H || W_idx < 0 || W_idx >= p.W) {
+                val = 0.0;
+            }
+            Bsh[B_ly * Bsh_stride + B_lx] = val;
+        }
+        barrier();
+        for (uint32_t CRS_lidx = 0; CRS_lidx < BS_CRS; CRS_lidx++) {
+            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                regA[T_ly] = Ash[(T_y * TS_K + T_ly) * Ash_stride + CRS_lidx];
+            }
+            for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                regB[T_lx] = Bsh[CRS_lidx * Bsh_stride + T_x * TS_NPQ + T_lx];
+            }
+            for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+                for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+                    regC[T_ly][T_lx] = fma(regA[T_ly], regB[T_lx], regC[T_ly][T_lx]);
+                }
+            }
+        }
+        barrier();
+    }
+    /* Save C* */
+    for (uint32_t T_ly = 0; T_ly < TS_K; T_ly++) {
+        for (uint32_t T_lx = 0; T_lx < TS_NPQ; T_lx++) {
+            uint32_t K_idx   = B_idx_K * BS_K + T_y * TS_K + T_ly;
+            uint32_t NPQ_idx = B_idx_NPQ * BS_NPQ + T_x * TS_NPQ + T_lx;
+            uint32_t N_idx   = NPQ_idx / (p.OH * p.OW);
+            uint32_t OH_idx  = (NPQ_idx - N_idx * p.OH * p.OW) / p.OW;
+            uint32_t OW_idx  = NPQ_idx - N_idx * p.OH * p.OW - OH_idx * p.OW;
+            uint32_t dst_idx = OW_idx + OH_idx * p.nb1 + K_idx * p.nb2 + N_idx * p.nb3;
+            if (K_idx < K && NPQ_idx < NPQ) {
+                dst_data[dst_idx] = regC[T_ly][T_lx];
+            }
+        }
+    }
+}

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/copy_to_quant.comp

@@ -1,22 +1,26 @@
 #version 450
 
-#if RTE16
-#extension GL_EXT_spirv_intrinsics : enable
-spirv_execution_mode(capabilities = [4467], 4462, 16); // RoundingModeRTE, 16 bits
-#endif // RTE16
-
+#include "rte.comp"
 #include "types.comp"
-#include "generic_unary_head.comp"
 
-#if defined(DATA_A_IQ4_NL)
-// 16 invocations needed for init_iq4nl_shmem
-layout(local_size_x = 16, local_size_y = 1, local_size_z = 1) in;
+#if defined(SET_ROWS) && QUANT_K == 1
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 512;
 #else
-layout(local_size_x = 1, local_size_y = 1, local_size_z = 1) in;
+layout(local_size_x = 32, local_size_y = 1, local_size_z = 1) in;
+const uint BLOCK_SIZE = 32;
 #endif
 
 layout (binding = 0) readonly buffer S {float data_s[];};
+
+#if defined(SET_ROWS)
+#include "generic_binary_head.comp"
+layout (binding = 1) readonly buffer C {uvec2 data_i[];};
+layout (binding = 2) writeonly buffer Q {A_TYPE data_q[];};
+#else
+#include "generic_unary_head.comp"
 layout (binding = 1) writeonly buffer Q {A_TYPE data_q[];};
+#endif
 
 #if defined(DATA_A_Q4_0)
 void quantize(uint dst_idx, uint src_idx)
@@ -221,15 +225,56 @@ void quantize(uint dst_idx, uint src_idx)
 }
 #endif
 
+#if defined(DATA_A_F32) || defined(DATA_A_F16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(data_s[src_idx]);
+}
+#endif
+
+#if defined(DATA_A_BF16)
+void quantize(uint dst_idx, uint src_idx)
+{
+    data_q[dst_idx] = A_TYPE(fp32_to_bf16(data_s[src_idx]));
+}
+#endif
+
+#if defined(SET_ROWS)
+
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
     init_iq_shmem(gl_WorkGroupSize);
-    if (gl_LocalInvocationIndex.x != 0) {
+#endif
+
+    const uint idx = ((gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x) * BLOCK_SIZE + gl_LocalInvocationID.x) * QUANT_K;
+
+    if (idx >= p.ne) {
         return;
     }
+
+    uint i00, i01, i02, i03;
+    get_indices(idx, i00, i01, i02, i03);
+
+    uint i12 = fastmod(i03, p.ne12);
+    uint i11 = fastmod(i02, p.ne11);
+    uint i10 = i01;
+
+    uint i1 = data_i[src1_idx(i10, i11, i12, 0) + get_boffset()].x;
+
+    uint src0_idx = src0_idx(i00, i01, i02, i03) + get_aoffset();
+    uint dst_idx = dst_idx(i00 / QUANT_K, i1, i02, i03) + get_doffset();
+
+    quantize(dst_idx, src0_idx);
+}
+
+#else
+
+void main() {
+#ifdef NEEDS_INIT_IQ_SHMEM
+    init_iq_shmem(gl_WorkGroupSize);
 #endif
 
-    const uint idx = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * QUANT_K;
+    const uint idx = (gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x * 32 + gl_LocalInvocationID.x) * QUANT_K;
 
     if (idx >= p.ne) {
         return;
@@ -240,3 +285,5 @@ void main() {
 
     quantize(dst_idx, src_idx);
 }
+
+#endif

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q2_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q3_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = uint(gl_WorkGroupID.x * 256 + wgy);
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q4_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q5_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint ib = gl_WorkGroupID.x * 256 + wgy;
-        if (ib >= p.M * p.K / QUANT_K) {
+        if (ib >= p.nel / QUANT_K) {
             return;
         }
 

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/dequant_q6_k.comp

@@ -10,7 +10,7 @@ layout (binding = 1) writeonly buffer D {D_TYPE data_b[];};
 void main() {
     [[unroll]] for (uint wgy = 0; wgy < 256; wgy++) {
         const uint i = gl_WorkGroupID.x * 256 + wgy;
-        if (i >= p.M * p.K / QUANT_K) {
+        if (i >= p.nel / QUANT_K) {
             return;
         }
         const uint tid = gl_LocalInvocationID.x;

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn.comp

@@ -11,7 +11,8 @@
 #include "types.comp"
 #include "flash_attn_base.comp"
 
-const uint32_t D_per_thread = D / D_split;
+const uint32_t HSK_per_thread = HSK / D_split;
+const uint32_t HSV_per_thread = HSV / D_split;
 
 const uint32_t cols_per_iter = WorkGroupSize / D_split;
 const uint32_t cols_per_thread = Bc / cols_per_iter;
@@ -29,7 +30,7 @@ layout (binding = 3) readonly buffer M {float16_t data_m[];};
 // Rows index by Q's dimension 2, and the first N rows are valid.
 D_TYPE perElemOpGqaStore(const in uint32_t r, const in uint32_t c, const in D_TYPE elem, const in uint32_t o_offset, const in uint32_t iq2, const in uint32_t N)
 {
-    uint32_t offset = (iq2 + r) * D + c;
+    uint32_t offset = (iq2 + r) * HSV + c;
     data_o[o_offset + offset] = D_TYPE(elem);
     return elem;
 }
@@ -38,7 +39,7 @@ shared FLOAT_TYPE tmpsh[WorkGroupSize];
 shared vec4 tmpshv4[WorkGroupSize];
 
 shared float masksh[Bc][Br];
-shared vec4 Qf[Br][D / 4];
+shared vec4 Qf[Br][HSK / 4];
 
 void main() {
 #ifdef NEEDS_INIT_IQ_SHMEM
@@ -53,18 +54,18 @@ void main() {
 
     uint32_t q_offset = (iq2*p.nb02+iq3*p.nb03) / 4;
 
-    [[unroll]] for (uint32_t idx = 0; idx < Br * D / 4; idx += gl_WorkGroupSize.x) {
-        uint32_t d = (idx + tid) % (D / 4);
-        uint32_t r = (idx + tid) / (D / 4);
-        if (r < Br && d < D / 4 &&
+    [[unroll]] for (uint32_t idx = 0; idx < Br * HSK / 4; idx += gl_WorkGroupSize.x) {
+        uint32_t d = (idx + tid) % (HSK / 4);
+        uint32_t r = (idx + tid) / (HSK / 4);
+        if (r < Br && d < HSK / 4 &&
             i * Br + r < N) {
             Qf[r][d] = vec4(data_qv4[q_offset / 4 + (i * Br + r) * q_stride / 4 + d]) * p.scale;
         }
     }
     barrier();
 
-    vec4 Of[Br][D_per_thread / 4];
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    vec4 Of[Br][HSV_per_thread / 4];
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] = vec4(0.0);
         }
@@ -99,6 +100,10 @@ void main() {
     uint32_t k_offset = (ik2*p.nb12 + ik3*p.nb13) / 2;
     uint32_t v_offset = (iv2*p.nb22 + iv3*p.nb23) / 2;
 #endif
+    uint32_t m_offset = 0;
+    if (p.nem2 != 1 || p.nem3 != 1) {
+        m_offset = ((iq3 % p.nem3) * p.nem2 + (iq2 % p.nem2)) * p.nem1 * KV;
+    }
 
     [[dont_unroll]]
     for (uint32_t j = start_j; j < end_j; ++j) {
@@ -112,7 +117,7 @@ void main() {
 
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSK_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * k_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -144,13 +149,13 @@ void main() {
             }
         }
 
-        if (p.mask != 0) {
+        if ((p.mask_n_head_log2 & MASK_ENABLE_BIT) != 0) {
 
             [[unroll]] for (uint32_t idx = 0; idx < Bc * Br; idx += gl_WorkGroupSize.x) {
                 uint32_t c = (idx + tid) % Bc;
                 uint32_t r = (idx + tid) / Bc;
                 if (idx + tid < Bc * Br) {
-                    masksh[c][r] = float(data_m[(i * Br + r) * m_stride + (j * Bc + c)]);
+                    masksh[c][r] = float(data_m[m_offset + (i * Br + r) * m_stride + (j * Bc + c)]);
                 }
             }
             barrier();
@@ -191,14 +196,14 @@ void main() {
             Lf[r] = eMf[r]*Lf[r] + rowsumf[r];
         }
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
             [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
                 Of[r][d] = eMf[r] * Of[r][d];
             }
         }
 
         [[unroll]] for (uint32_t c = 0; c < cols_per_thread; ++c) {
-            [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+            [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 #if BLOCK_SIZE > 1
                 uint coord = (j * Bc + c * cols_per_iter + col_tid) * v_stride * BLOCK_SIZE + 4 * (d * D_split + d_tid);
                 uint ib = coord / BLOCK_SIZE;
@@ -255,7 +260,7 @@ void main() {
         Lf[r] = tmpsh[d_tid];
         barrier();
 
-        [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+        [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
 
             Of[r][d] = eMf * Of[r][d];
             tmpshv4[tid] = Of[r][d];
@@ -277,11 +282,11 @@ void main() {
     // If there is split_k, then the split_k resolve shader does the final
     // division by L. Store the intermediate O value and per-row m and L values.
     if (p.k_num > 1) {
-        uint32_t o_offset = D * p.ne1 * split_k_index;
+        uint32_t o_offset = HSV * p.ne1 * (split_k_index + iq3 * p.k_num);
 
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -289,7 +294,7 @@ void main() {
             }
         }
 
-        o_offset = D * p.ne1 * p.k_num + p.ne1 * split_k_index * 2;
+        o_offset = HSV * p.ne1 * p.ne3 * p.k_num + p.ne1 * (split_k_index + iq3 * p.k_num) * 2;
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
                 perElemOpStoreCol0(r, 0u, ACC_TYPE(Lf[r]), o_offset, iq2, N);
@@ -305,18 +310,18 @@ void main() {
         Lfrcp[r] = 1.0 / Lf[r];
     }
 
-    [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+    [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             Of[r][d] *= Lfrcp[r];
         }
     }
 
-    uint32_t o_offset = iq3*p.ne2*p.ne1;
+    uint32_t o_offset = iq3*p.ne2*p.ne1*HSV;
 
     if (p.gqa_ratio > 1) {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
                         perElemOpGqaStore(r, 4*(d * D_split + d_tid) + comp, Of[r][d][comp], o_offset, iq2, N);
                     }
@@ -326,9 +331,9 @@ void main() {
     } else {
         [[unroll]] for (uint32_t r = 0; r < Br; ++r) {
             if (i * Br + r < N) {
-                [[unroll]] for (uint32_t d = 0; d < D_per_thread / 4; ++d) {
+                [[unroll]] for (uint32_t d = 0; d < HSV_per_thread / 4; ++d) {
                     [[unroll]] for (uint32_t comp = 0; comp < 4; ++comp) {
-                        data_o[o_offset + iq2 * D + (i * Br + r) * p.ne1 * D + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
+                        data_o[o_offset + iq2 * HSV + (i * Br + r) * p.ne1 * HSV + 4*(d * D_split + d_tid) + comp] = D_TYPE(Of[r][d][comp]);
                     }
                 }
             }

package/cpp/llama.cpp/ggml/src/ggml-vulkan/vulkan-shaders/flash_attn_base.comp

@@ -4,10 +4,10 @@ layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
 layout (constant_id = 0) const uint32_t WorkGroupSize = 128;
 layout (constant_id = 1) const uint32_t Br = 1;
 layout (constant_id = 2) const uint32_t Bc = 32;
-layout (constant_id = 3) const uint32_t D = 32;
-layout (constant_id = 4) const uint32_t Clamp = 0;
-layout (constant_id = 5) const uint32_t D_split = 16;
-
+layout (constant_id = 3) const uint32_t HSK = 32;
+layout (constant_id = 4) const uint32_t HSV = 32;
+layout (constant_id = 5) const uint32_t Clamp = 0;
+layout (constant_id = 6) const uint32_t D_split = 16;
 
 layout (push_constant) uniform parameter {
     uint32_t N;
@@ -24,6 +24,8 @@ layout (push_constant) uniform parameter {
     uint32_t nev2;
     uint32_t nev3;
     uint32_t nem1;
+    uint32_t nem2;
+    uint32_t nem3;
 
     uint32_t nb01;
     uint32_t nb02;
@@ -34,14 +36,12 @@ layout (push_constant) uniform parameter {
     uint32_t nb21;
     uint32_t nb22;
     uint32_t nb23;
-    uint32_t nb31;
 
     float scale;
     float max_bias;
     float logit_softcap;
 
-    uint32_t mask;
-    uint32_t n_head_log2;
+    uint32_t mask_n_head_log2;
     float m0;
     float m1;
 
@@ -50,6 +50,9 @@ layout (push_constant) uniform parameter {
     uint32_t k_num;
 } p;
 
+#define MASK_ENABLE_BIT (1<<16)
+#define N_LOG2_MASK 0xFFFF
+
 layout (binding = 4) writeonly buffer O {D_TYPE data_o[];};
 
 #if defined(A_TYPE_PACKED16)
@@ -100,8 +103,10 @@ ACC_TYPE perElemOpComputeSlope(const in uint32_t r, const in uint32_t c, const i
 {
     const uint32_t h = iq2 + (r % p.gqa_ratio);
 
-    const ACC_TYPE base = ACC_TYPE(h < p.n_head_log2 ? p.m0 : p.m1);
-    const int      exph = int(h < p.n_head_log2 ? h + 1 : 2*(h - p.n_head_log2) + 1);
+    uint32_t n_head_log2 = p.mask_n_head_log2 & N_LOG2_MASK;
+
+    const ACC_TYPE base = ACC_TYPE(h < n_head_log2 ? p.m0 : p.m1);
+    const int      exph = int(h < n_head_log2 ? h + 1 : 2*(h - n_head_log2) + 1);
 
     return ACC_TYPE(pow(base, ACC_TYPE(exph)));
 }