mirror of
https://github.com/LostRuins/koboldcpp.git
synced 2025-09-11 09:34:37 +00:00
Merge branch 'upstream' into concedo_experimental
# Conflicts: # .github/workflows/build.yml # README.md # examples/CMakeLists.txt # examples/batched/batched.cpp # examples/gritlm/gritlm.cpp # examples/llama.android/llama/build.gradle.kts # examples/main/README.md # examples/retrieval/retrieval.cpp # examples/server/CMakeLists.txt # examples/server/README.md # ggml/CMakeLists.txt # ggml/src/ggml-cpu/CMakeLists.txt # ggml/src/ggml.c # scripts/compare-commits.sh # scripts/sync-ggml.last # tests/CMakeLists.txt # tests/test-backend-ops.cpp # tests/test-chat-template.cpp # tests/test-sampling.cpp
This commit is contained in:
Concedo
commit
ee486bad3e
59 changed files with 20531 additions and 13185 deletions
|
|
@ -245,6 +245,7 @@ struct vk_device_struct {
|
|||
vk_pipeline pipeline_im2col_f32, pipeline_im2col_f32_f16;
|
||||
vk_pipeline pipeline_timestep_embedding_f32;
|
||||
vk_pipeline pipeline_pool2d_f32;
|
||||
vk_pipeline pipeline_rwkv_wkv6_f32;
|
||||
|
||||
// [2][2][2] is for {f16acc,f32acc}x{large,small_rows}x{unaligned, aligned}
|
||||
vk_pipeline pipeline_flash_attn_f32_f16_D64[GGML_TYPE_COUNT][2][2][2];
|
||||
|
|
@ -528,6 +529,13 @@ struct vk_op_pool2d_push_constants {
|
|||
int32_t p0; int32_t p1;
|
||||
};
|
||||
|
||||
struct vk_op_rwkv_wkv6_push_constants {
|
||||
uint32_t B;
|
||||
uint32_t T;
|
||||
uint32_t C;
|
||||
uint32_t H;
|
||||
};
|
||||
|
||||
// Allow pre-recording command buffers
|
||||
struct vk_staging_memcpy {
|
||||
vk_staging_memcpy(void * _dst, const void * _src, size_t _n) : dst(_dst), src(_src), n(_n) {}
|
||||
|
|
@ -1363,7 +1371,7 @@ static bool ggml_vk_matmul_shmem_support(const vk_device& device, const std::vec
|
|||
// Needs to be kept up to date on shader changes
|
||||
const uint32_t bank_conflict_offset = device->coopmat_support ? 8 : 1;
|
||||
const uint32_t type_size = device->fp16 ? sizeof(ggml_fp16_t) : sizeof(float);
|
||||
const uint32_t warps = warptile[0] / device->subgroup_size;
|
||||
const uint32_t warps = warptile[0] / warptile[10];
|
||||
|
||||
const uint32_t load_bufs = (warptile[1] + warptile[2]) * (warptile[3] + bank_conflict_offset) * type_size;
|
||||
const uint32_t mmid_row_ids = mul_mat_id ? 3072 * sizeof(uint32_t) : 0;
|
||||
|
|
@ -1377,8 +1385,9 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
std::cerr << "ggml_vulkan: Compiling shaders";
|
||||
|
||||
// some shaders require the subgroup size to be 16 or larger
|
||||
// some shaders have a minimum subgroup size
|
||||
const uint32_t subgroup_size_16 = std::max(device->subgroup_size, 16u);
|
||||
const uint32_t subgroup_size_32 = std::max(device->subgroup_size, 32u);
|
||||
|
||||
// mulmat
|
||||
std::vector<uint32_t> l_warptile, m_warptile, s_warptile,
|
||||
|
|
@ -1445,7 +1454,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
l_warptile_mmq = { 128, 128, 128, 32, device->subgroup_size * 2, 64, 2, tm_l, tn_l, tk_l, device->subgroup_size };
|
||||
m_warptile_mmq = { 128, 64, 64, 32, device->subgroup_size, 32, 2, tm_m, tn_m, tk_m, device->subgroup_size };
|
||||
s_warptile_mmq = { subgroup_size_16, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
|
||||
s_warptile_mmq = { subgroup_size_32, 32, 32, 32, 32, 32, 2, tm_s, tn_s, tk_s, device->subgroup_size };
|
||||
|
||||
l_mmq_wg_denoms = l_wg_denoms = {128, 128, 1 };
|
||||
m_mmq_wg_denoms = m_wg_denoms = { 64, 64, 1 };
|
||||
|
|
@ -1864,7 +1873,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f32_f32", mul_mat_vec_q4_k_f32_f32_len, mul_mat_vec_q4_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f32_f32", mul_mat_vec_q5_k_f32_f32_len, mul_mat_vec_q5_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f32_f32", mul_mat_vec_q6_k_f32_f32_len, mul_mat_vec_q6_k_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f32_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f32_f32", mul_mat_vec_iq4_nl_f32_f32_len, mul_mat_vec_iq4_nl_f32_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
|
||||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F32 ], "mul_mat_vec_f32_f16_f32", mul_mat_vec_f32_f16_f32_len, mul_mat_vec_f32_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_F16 ], "mul_mat_vec_f16_f16_f32", mul_mat_vec_f16_f16_f32_len, mul_mat_vec_f16_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
|
||||
|
|
@ -1878,7 +1887,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q4_K], "mul_mat_vec_q4_k_f16_f32", mul_mat_vec_q4_k_f16_f32_len, mul_mat_vec_q4_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q5_K], "mul_mat_vec_q5_k_f16_f32", mul_mat_vec_q5_k_f16_f32_len, mul_mat_vec_q5_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_Q6_K], "mul_mat_vec_q6_k_f16_f32", mul_mat_vec_q6_k_f16_f32_len, mul_mat_vec_q6_k_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_f16_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_iq4_nl_f16_f32", mul_mat_vec_iq4_nl_f16_f32_len, mul_mat_vec_iq4_nl_f16_f32_data, "main", 3, sizeof(vk_mat_vec_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
|
||||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F32 ], "mul_mat_vec_id_f32_f32", mul_mat_vec_id_f32_f32_len, mul_mat_vec_id_f32_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_F16 ], "mul_mat_vec_id_f16_f32", mul_mat_vec_id_f16_f32_len, mul_mat_vec_id_f16_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2, 1, 1}, {device->subgroup_size, 2}, 1);
|
||||
|
|
@ -1892,7 +1901,7 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q4_K], "mul_mat_vec_id_q4_k_f32", mul_mat_vec_id_q4_k_f32_len, mul_mat_vec_id_q4_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q5_K], "mul_mat_vec_id_q5_k_f32", mul_mat_vec_id_q5_k_f32_len, mul_mat_vec_id_q5_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_Q6_K], "mul_mat_vec_id_q6_k_f32", mul_mat_vec_id_q6_k_f32_len, mul_mat_vec_id_q6_k_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {1, 1, 1}, {subgroup_size_16}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {device->subgroup_size, 2*rm}, 1, true);
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant_mul_mat_vec_id_f32[GGML_TYPE_IQ4_NL], "mul_mat_vec_id_iq4_nl_f32", mul_mat_vec_id_iq4_nl_f32_len, mul_mat_vec_id_iq4_nl_f32_data, "main", 4, sizeof(vk_mat_vec_id_push_constants), {2*rm, 1, 1}, {subgroup_size_16, 2*rm}, 1, true);
|
||||
|
||||
// dequant shaders
|
||||
ggml_vk_create_pipeline(device, device->pipeline_dequant[GGML_TYPE_F32 ], "f32_to_f16", dequant_f32_len, dequant_f32_data, "main", 2, 5 * sizeof(uint32_t), {256 * 16, 1, 1}, {}, 1);
|
||||
|
|
@ -2014,6 +2023,8 @@ static void ggml_vk_load_shaders(vk_device& device) {
|
|||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_pool2d_f32, "pool2d_f32", pool2d_f32_len, pool2d_f32_data, "main", 2, sizeof(vk_op_pool2d_push_constants), {512, 1, 1}, {}, 1);
|
||||
|
||||
ggml_vk_create_pipeline(device, device->pipeline_rwkv_wkv6_f32, "rwkv_wkv6_f32", rwkv_wkv6_f32_len, rwkv_wkv6_f32_data, "main", 7, sizeof(vk_op_rwkv_wkv6_push_constants), {1, 1, 1}, {device->subgroup_size}, 1);
|
||||
|
||||
for (auto &c : compiles) {
|
||||
c.wait();
|
||||
}
|
||||
|
|
@ -5022,6 +5033,11 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
|
|||
return ctx->device->pipeline_pool2d_f32;
|
||||
}
|
||||
return nullptr;
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
||||
return ctx->device->pipeline_rwkv_wkv6_f32;
|
||||
}
|
||||
return nullptr;
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {
|
||||
return ctx->device->pipeline_leaky_relu_f32;
|
||||
|
|
@ -5424,6 +5440,134 @@ static void ggml_vk_div(ggml_backend_vk_context * ctx, vk_context& subctx, const
|
|||
}, dryrun);
|
||||
}
|
||||
|
||||
static void ggml_vk_op_f32_rwkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, const vk_op_rwkv_wkv6_push_constants&& pc, bool dryrun = false) {
|
||||
const ggml_tensor * k = dst->src[0];
|
||||
const ggml_tensor * v = dst->src[1];
|
||||
const ggml_tensor * r = dst->src[2];
|
||||
const ggml_tensor * tf = dst->src[3];
|
||||
const ggml_tensor * td = dst->src[4];
|
||||
const ggml_tensor * state = dst->src[5];
|
||||
|
||||
GGML_ASSERT(!ggml_is_quantized(k->type));
|
||||
GGML_ASSERT(!ggml_is_quantized(v->type));
|
||||
GGML_ASSERT(!ggml_is_quantized(r->type));
|
||||
GGML_ASSERT(!ggml_is_quantized(tf->type));
|
||||
GGML_ASSERT(!ggml_is_quantized(td->type));
|
||||
GGML_ASSERT(!ggml_is_quantized(state->type));
|
||||
GGML_ASSERT(dst->buffer != nullptr);
|
||||
|
||||
vk_pipeline pipeline = ggml_vk_op_get_pipeline(ctx, k, v, r, dst, GGML_OP_RWKV_WKV6);
|
||||
GGML_ASSERT(pipeline != nullptr);
|
||||
|
||||
if (dryrun) {
|
||||
ggml_pipeline_request_descriptor_sets(ctx->device, pipeline, 1);
|
||||
return;
|
||||
}
|
||||
|
||||
ggml_backend_vk_buffer_context * dst_buf_ctx = (ggml_backend_vk_buffer_context *)dst->buffer->context;
|
||||
ggml_backend_vk_buffer_context * k_buf_ctx = (ggml_backend_vk_buffer_context *)k->buffer->context;
|
||||
ggml_backend_vk_buffer_context * v_buf_ctx = (ggml_backend_vk_buffer_context *)v->buffer->context;
|
||||
ggml_backend_vk_buffer_context * r_buf_ctx = (ggml_backend_vk_buffer_context *)r->buffer->context;
|
||||
ggml_backend_vk_buffer_context * tf_buf_ctx = (ggml_backend_vk_buffer_context *)tf->buffer->context;
|
||||
ggml_backend_vk_buffer_context * td_buf_ctx = (ggml_backend_vk_buffer_context *)td->buffer->context;
|
||||
ggml_backend_vk_buffer_context * state_buf_ctx = (ggml_backend_vk_buffer_context *)state->buffer->context;
|
||||
|
||||
ggml_vk_sync_buffers(subctx);
|
||||
|
||||
vk_buffer d_D, d_K, d_V, d_R, d_TF, d_TD, d_State;
|
||||
uint64_t k_offset, v_offset, r_offset, tf_offset, td_offset, state_offset, dst_offset;
|
||||
bool K_uma = false, V_uma = false, R_uma = false, TF_uma = false, TD_uma = false, STATE_uma = false, DST_uma = false;
|
||||
|
||||
if (ctx->device->uma) {
|
||||
ggml_vk_host_get(ctx->device, k->data, d_K, k_offset);
|
||||
ggml_vk_host_get(ctx->device, v->data, d_V, v_offset);
|
||||
ggml_vk_host_get(ctx->device, r->data, d_R, r_offset);
|
||||
ggml_vk_host_get(ctx->device, tf->data, d_TF, tf_offset);
|
||||
ggml_vk_host_get(ctx->device, td->data, d_TD, td_offset);
|
||||
ggml_vk_host_get(ctx->device, state->data, d_State, state_offset);
|
||||
ggml_vk_host_get(ctx->device, dst->data, d_D, dst_offset);
|
||||
|
||||
K_uma = d_K != nullptr;
|
||||
V_uma = d_V != nullptr;
|
||||
R_uma = d_R != nullptr;
|
||||
TF_uma = d_TF != nullptr;
|
||||
TD_uma = d_TD != nullptr;
|
||||
STATE_uma = d_State != nullptr;
|
||||
DST_uma = d_D != nullptr;
|
||||
}
|
||||
|
||||
if (!K_uma) {
|
||||
d_K = k_buf_ctx->dev_buffer;
|
||||
k_offset = vk_tensor_offset(k) + k->view_offs;
|
||||
}
|
||||
if (!V_uma) {
|
||||
d_V = v_buf_ctx->dev_buffer;
|
||||
v_offset = vk_tensor_offset(v) + v->view_offs;
|
||||
}
|
||||
if (!R_uma) {
|
||||
d_R = r_buf_ctx->dev_buffer;
|
||||
r_offset = vk_tensor_offset(r) + r->view_offs;
|
||||
}
|
||||
if (!TF_uma) {
|
||||
d_TF = tf_buf_ctx->dev_buffer;
|
||||
tf_offset = vk_tensor_offset(tf) + tf->view_offs;
|
||||
}
|
||||
if (!TD_uma) {
|
||||
d_TD = td_buf_ctx->dev_buffer;
|
||||
td_offset = vk_tensor_offset(td) + td->view_offs;
|
||||
}
|
||||
if (!STATE_uma) {
|
||||
d_State = state_buf_ctx->dev_buffer;
|
||||
state_offset = vk_tensor_offset(state) + state->view_offs;
|
||||
}
|
||||
if (!DST_uma) {
|
||||
d_D = dst_buf_ctx->dev_buffer;
|
||||
dst_offset = vk_tensor_offset(dst) + dst->view_offs;
|
||||
}
|
||||
|
||||
const uint64_t k_size = ggml_nbytes(k);
|
||||
const uint64_t v_size = ggml_nbytes(v);
|
||||
const uint64_t r_size = ggml_nbytes(r);
|
||||
const uint64_t tf_size = ggml_nbytes(tf);
|
||||
const uint64_t td_size = ggml_nbytes(td);
|
||||
const uint64_t state_size = ggml_nbytes(state);
|
||||
const uint64_t dst_size = ggml_nbytes(dst);
|
||||
|
||||
std::array<uint32_t, 3> elements = {
|
||||
(uint32_t)(pc.B * pc.H),
|
||||
1,
|
||||
1
|
||||
};
|
||||
|
||||
ggml_vk_dispatch_pipeline(ctx, subctx, pipeline, {
|
||||
vk_subbuffer{ d_K, k_offset, k_size },
|
||||
vk_subbuffer{ d_V, v_offset, v_size },
|
||||
vk_subbuffer{ d_R, r_offset, r_size },
|
||||
vk_subbuffer{ d_TF, tf_offset, tf_size },
|
||||
vk_subbuffer{ d_TD, td_offset, td_size },
|
||||
vk_subbuffer{ d_State, state_offset, state_size },
|
||||
vk_subbuffer{ d_D, dst_offset, dst_size }
|
||||
}, sizeof(vk_op_rwkv_wkv6_push_constants), &pc, elements);
|
||||
}
|
||||
|
||||
static void ggml_vk_rwkv_wkv6(ggml_backend_vk_context * ctx, vk_context& subctx, ggml_tensor * dst, bool dryrun = false) {
|
||||
const size_t seq_length = dst->src[0]->ne[3];
|
||||
const size_t n_embed = dst->ne[0];
|
||||
const size_t n_heads = dst->src[0]->ne[2];
|
||||
const size_t n_seqs = dst->src[5]->ne[1];
|
||||
|
||||
ggml_vk_op_f32_rwkv6(
|
||||
ctx, subctx, dst,
|
||||
{
|
||||
(uint32_t)n_seqs,
|
||||
(uint32_t)seq_length,
|
||||
(uint32_t)n_embed,
|
||||
(uint32_t)n_heads,
|
||||
},
|
||||
dryrun
|
||||
);
|
||||
}
|
||||
|
||||
static void ggml_vk_concat(ggml_backend_vk_context * ctx, vk_context& subctx, const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst, bool dryrun = false) {
|
||||
int * op_params = (int *)dst->op_params;
|
||||
|
||||
|
|
@ -6569,6 +6713,7 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
case GGML_OP_IM2COL:
|
||||
case GGML_OP_TIMESTEP_EMBEDDING:
|
||||
case GGML_OP_POOL_2D:
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
case GGML_OP_FLASH_ATTN_EXT:
|
||||
break;
|
||||
|
|
@ -6768,6 +6913,11 @@ static bool ggml_vk_build_graph(ggml_backend_vk_context * ctx, ggml_tensor * nod
|
|||
case GGML_OP_FLASH_ATTN_EXT:
|
||||
ggml_vk_flash_attn(ctx, compute_ctx, src0, src1, src2, src3, node, dryrun);
|
||||
|
||||
break;
|
||||
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
ggml_vk_rwkv_wkv6(ctx, compute_ctx, node, dryrun);
|
||||
|
||||
break;
|
||||
default:
|
||||
return false;
|
||||
|
|
@ -6848,6 +6998,7 @@ static bool ggml_vk_compute_forward(ggml_backend_vk_context * ctx, ggml_tensor *
|
|||
case GGML_OP_IM2COL:
|
||||
case GGML_OP_TIMESTEP_EMBEDDING:
|
||||
case GGML_OP_POOL_2D:
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
case GGML_OP_REPEAT:
|
||||
buf = tensor->buffer;
|
||||
|
|
@ -7724,6 +7875,7 @@ static bool ggml_backend_vk_device_supports_op(ggml_backend_dev_t dev, const ggm
|
|||
case GGML_OP_IM2COL:
|
||||
case GGML_OP_TIMESTEP_EMBEDDING:
|
||||
case GGML_OP_POOL_2D:
|
||||
case GGML_OP_RWKV_WKV6:
|
||||
case GGML_OP_LEAKY_RELU:
|
||||
return true;
|
||||
default:
|
||||
|
|
@ -8300,7 +8452,11 @@ static void ggml_vk_check_results_0(ggml_tensor * tensor) {
|
|||
} else if (tensor->op == GGML_OP_LEAKY_RELU) {
|
||||
const float * op_params = (const float *)tensor->op_params;
|
||||
tensor_clone = ggml_leaky_relu(ggml_ctx, src0_clone, op_params[0], false);
|
||||
} else {
|
||||
} else if (tensor->op == GGML_OP_RWKV_WKV6) {
|
||||
tensor_clone = ggml_rwkv_wkv6(ggml_ctx, tensor->src[0], tensor->src[1], tensor->src[2], tensor->src[3],
|
||||
tensor->src[4], tensor->src[5]);
|
||||
}
|
||||
else {
|
||||
std::cerr << "Missing vk_check_results OP: " << ggml_op_name(tensor->op) << std::endl;
|
||||
GGML_ABORT("fatal error");
|
||||
}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue