@fugood/llama.node 0.3.14 → 0.3.15

package/src/llama.cpp/tests/test-backend-ops.cpp

@@ -259,6 +259,10 @@ static std::string var_to_str(ggml_type type) {
     return ggml_type_name(type);
 }
 
+static std::string var_to_str(ggml_prec prec) {
+    return prec == GGML_PREC_F32 ? "f32" : "def";
+}
+
 static std::string var_to_str(ggml_op_pool pool) {
     switch (pool) {
         case GGML_OP_POOL_AVG:  return "avg";
@@ -1916,6 +1920,40 @@ struct test_gla : public test_case {
     }
 };
 
+// GGML_OP_RWKV_WKV7
+struct test_rwkv_wkv7 : public test_case {
+    const ggml_type type;
+
+    const int64_t head_count;
+    const int64_t head_size;
+    const int64_t n_seq_tokens;
+    const int64_t n_seqs;
+
+    std::string vars() override {
+        return VARS_TO_STR5(type, head_count, head_size, n_seq_tokens, n_seqs);
+    }
+
+    test_rwkv_wkv7(ggml_type type = GGML_TYPE_F32,
+            int64_t head_count = 32, int64_t head_size = 64, int64_t n_seq_tokens = 32, int64_t n_seqs = 32)
+        : type(type), head_count(head_count), head_size(head_size), n_seq_tokens(n_seq_tokens), n_seqs(n_seqs) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        const int64_t n_tokens = n_seq_tokens * n_seqs;
+        ggml_tensor * r   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * w   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * k   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * v   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * a   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        ggml_tensor * b   = ggml_new_tensor(ctx, type, 3, std::vector<int64_t>{ head_size, head_count, n_tokens }.data());
+        // Outputs may become NaN with long seqlen without these normalization
+        a = ggml_l2_norm(ctx, a, 1e-7F);
+        b = ggml_l2_norm(ctx, b, 1e-7F);
+        ggml_tensor * s   = ggml_new_tensor(ctx, type, 2, std::vector<int64_t>{ head_size * head_size * head_count, n_seqs }.data());
+        ggml_tensor * out = ggml_rwkv_wkv7(ctx, r, w, k, v, a, b, s);
+        return out;
+    }
+};
+
 // GGML_OP_MUL_MAT
 struct test_mul_mat : public test_case {
     const ggml_type type_a;
@@ -2972,6 +3010,32 @@ struct test_group_norm : public test_case {
     }
 };
 
+// GGML_OP_L2_NORM
+struct test_l2_norm : public test_case {
+    const ggml_type type;
+    const std::array<int64_t, 4> ne;
+    const float eps;
+
+    std::string vars() override {
+        return VARS_TO_STR2(type, ne);
+    }
+
+    test_l2_norm(ggml_type type = GGML_TYPE_F32,
+            std::array<int64_t, 4> ne = {64, 64, 320, 1},
+            float eps = 1e-12f)
+        : type(type), ne(ne), eps(eps) {}
+
+    ggml_tensor * build_graph(ggml_context * ctx) override {
+        ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());
+        ggml_set_name(a, "a");
+
+        ggml_tensor * out = ggml_l2_norm(ctx, a, eps);
+        ggml_set_name(out, "out");
+
+        return out;
+    }
+};
+
 // GGML_OP_ACC
 struct test_acc : public test_case {
     const ggml_type type;
@@ -3146,11 +3210,12 @@ struct test_flash_attn_ext : public test_case {
     const float max_bias; // ALiBi
     const float logit_softcap; // Gemma 2
 
+    const ggml_prec prec;
     const ggml_type type_KV;
     std::array<int32_t, 4> permute;
 
     std::string vars() override {
-        return VARS_TO_STR10(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, permute);
+        return VARS_TO_STR11(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, permute);
     }
 
     double max_nmse_err() override {
@@ -3165,9 +3230,9 @@ struct test_flash_attn_ext : public test_case {
     }
 
     test_flash_attn_ext(int64_t hs = 128, int64_t nh = 32, int64_t nr = 1, int64_t kv = 96, int64_t nb = 8,
-                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_type type_KV = GGML_TYPE_F16,
-                        std::array<int32_t, 4> permute = {0, 1, 2, 3})
-        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), type_KV(type_KV), permute(permute) {}
+                        bool mask = true, float max_bias = 0.0f, float logit_softcap = 0.0f, ggml_prec prec = GGML_PREC_F32,
+                        ggml_type type_KV = GGML_TYPE_F16, std::array<int32_t, 4> permute = {0, 1, 2, 3})
+        : hs(hs), nh(nh), nr(nr), kv(kv), nb(nb), mask(mask), max_bias(max_bias), logit_softcap(logit_softcap), prec(prec), type_KV(type_KV), permute(permute) {}
 
     ggml_tensor * build_graph(ggml_context * ctx) override {
         const int64_t hs_padded = GGML_PAD(hs, ggml_blck_size(type_KV));
@@ -3201,6 +3266,7 @@ struct test_flash_attn_ext : public test_case {
         }
 
         ggml_tensor * out = ggml_flash_attn_ext(ctx, q, k, v, m, 1.0f/sqrtf(hs), max_bias, logit_softcap);
+        ggml_flash_attn_ext_set_prec(out, prec);
         ggml_set_name(out, "out");
 
         return out;
@@ -4006,8 +4072,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
             test_cases.emplace_back(new test_rms_norm(GGML_TYPE_F32, {64, 5, 4, 3}, v, eps));
         }
         test_cases.emplace_back(new test_rms_norm_back(GGML_TYPE_F32, {64, 5, 4, 3}, eps));
+        test_cases.emplace_back(new test_l2_norm      (GGML_TYPE_F32, {64, 5, 4, 3}, eps));
     }
 
+    test_cases.emplace_back(new test_l2_norm(GGML_TYPE_F32, {64, 5, 4, 3}, 1e-12f));
+
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 1, 1}, {4, 1536, 1, 1}));
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {8, 1536, 1, 1}, {4, 1536, 1, 1}));
     test_cases.emplace_back(new test_ssm_conv(GGML_TYPE_F32, {4, 1536, 4, 1}, {4, 1536, 1, 1}));
@@ -4019,6 +4088,11 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 32, 4));
     test_cases.emplace_back(new test_rwkv_wkv6(GGML_TYPE_F32, 32, 64, 128, 4));
 
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 1, 1));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 1));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 32, 4));
+    test_cases.emplace_back(new test_rwkv_wkv7(GGML_TYPE_F32, 32, 64, 128, 4));
+
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 1, 1));
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 1));
     test_cases.emplace_back(new test_gla(GGML_TYPE_F32, 32, 64, 32, 4));
@@ -4308,11 +4382,16 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
                             for (int kv : { 512, 1024, }) {
                                 if (nr != 1 && kv != 512) continue;
                                 for (int nb : { 1, 3, 32, 35, }) {
-                                    for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
-                                        test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV));
-                                        // run fewer test cases permuted
-                                        if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
-                                            test_cases.emplace_back(new test_flash_attn_ext(hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, type_KV, {0, 2, 1, 3}));
+                                    for (ggml_prec prec : {GGML_PREC_F32, GGML_PREC_DEFAULT}) {
+                                        if (hs != 128 && prec == GGML_PREC_DEFAULT) continue;
+                                        for (ggml_type type_KV : {GGML_TYPE_F16, GGML_TYPE_BF16, GGML_TYPE_Q8_0, GGML_TYPE_Q4_0}) {
+                                            test_cases.emplace_back(new test_flash_attn_ext(
+                                                hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV));
+                                            // run fewer test cases permuted
+                                            if (mask == true && max_bias == 0.0f && logit_softcap == 0 && kv == 512) {
+                                                test_cases.emplace_back(new test_flash_attn_ext(
+                                                    hs, nh, nr, kv, nb, mask, max_bias, logit_softcap, prec, type_KV, {0, 2, 1, 3}));
+                                            }
                                         }
                                     }
                                 }

package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.cpp

@@ -1,143 +0,0 @@
-#include <sycl/sycl.hpp>
-#include "wkv6.hpp"
-
-constexpr int WKV_BLOCK_SIZE = 64;  // Matching CUDA_WKV_BLOCK_SIZE
-
-// Helper function for the main kernel
-static void rwkv_wkv_f32_kernel(
-    const int B, const int T, const int C, const int H,
-    const float* k, const float* v, const float* r,
-    const float* tf, const float* td, const float* s,
-    float* dst, const sycl::nd_item<3>& item_ct1, float* shared_mem) {
-
-    const int tid = item_ct1.get_local_id(2);
-    const int bid = item_ct1.get_group(2);
-
-    const int head_size = WKV_BLOCK_SIZE;
-    const int batch_i = bid / H;
-    const int head_i = bid % H;
-    const int state_size = C * head_size;
-    const int n_seq_tokens = T / B;
-
-    // Set up shared memory pointers
-    float* _k = shared_mem;
-    float* _r = _k + head_size;
-    float* _tf = _r + head_size;
-    float* _td = _tf + head_size;
-
-    // Local state array
-    float state[WKV_BLOCK_SIZE];
-
-    // Load initial state
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        state[i] = s[batch_i * state_size + head_i * head_size * head_size + i * head_size + tid];
-    }
-
-    // Sync threads before shared memory operations
-    item_ct1.barrier(sycl::access::fence_space::local_space);
-
-    // Load time-mixing parameters
-    _tf[tid] = tf[head_i * head_size + tid];
-    item_ct1.barrier(sycl::access::fence_space::local_space);
-
-    // Main sequence processing loop
-    for (int t = batch_i * n_seq_tokens * C + head_i * head_size + tid;
-         t < (batch_i + 1) * n_seq_tokens * C + head_i * head_size + tid;
-         t += C) {
-
-        item_ct1.barrier(sycl::access::fence_space::local_space);
-
-        // Load current timestep data to shared memory
-        _k[tid] = k[t];
-        _r[tid] = r[t];
-        _td[tid] = td[t];
-
-        item_ct1.barrier(sycl::access::fence_space::local_space);
-
-        const float _v = v[t];
-        float y = 0;
-
-        // Process in chunks of 4 for better vectorization
-        sycl::float4 k4, r4, tf4, td4, s4;
-        #pragma unroll
-        for (int j = 0; j < head_size; j += 4) {
-            // Load data in vec4 chunks
-            k4 = sycl::float4(_k[j], _k[j+1], _k[j+2], _k[j+3]);
-            r4 = sycl::float4(_r[j], _r[j+1], _r[j+2], _r[j+3]);
-            tf4 = sycl::float4(_tf[j], _tf[j+1], _tf[j+2], _tf[j+3]);
-            td4 = sycl::float4(_td[j], _td[j+1], _td[j+2], _td[j+3]);
-            s4 = sycl::float4(state[j], state[j+1], state[j+2], state[j+3]);
-
-            // Compute key-value product
-            sycl::float4 kv4 = k4 * _v;
-
-            // Accumulate weighted sum
-            y += sycl::dot(r4, tf4 * kv4 + s4);
-
-            // Update state
-            s4 = s4 * td4 + kv4;
-
-            // Store updated state
-            state[j] = s4.x();
-            state[j+1] = s4.y();
-            state[j+2] = s4.z();
-            state[j+3] = s4.w();
-        }
-
-        dst[t] = y;
-    }
-
-    // Save final state
-    #pragma unroll
-    for (int i = 0; i < head_size; i++) {
-        dst[T * C + batch_i * state_size + head_i * head_size * head_size + i * head_size + tid] = state[i];
-    }
-}
-
-void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
-
-    const ggml_tensor *src0 = dst->src[0];
-    const ggml_tensor *src1 = dst->src[1];
-
-    const float* k_d = (const float*)dst->src[0]->data;
-    const float* v_d = (const float*)dst->src[1]->data;
-    const float* r_d = (const float*)dst->src[2]->data;
-    const float* tf_d = (const float*)dst->src[3]->data;
-    const float* td_d = (const float*)dst->src[4]->data;
-    const float* s_d = (const float*)dst->src[5]->data;
-    float* dst_d = (float*)dst->data;
-
-    const int64_t B = dst->src[5]->ne[1];
-    const int64_t T = dst->src[0]->ne[2];
-    const int64_t C = dst->ne[0];
-    const int64_t H = dst->src[0]->ne[1];
-
-    GGML_ASSERT(dst->src[5]->type == GGML_TYPE_F32);
-    GGML_ASSERT(C % H == 0);
-    GGML_ASSERT(C / H == WKV_BLOCK_SIZE); // The current sycl kernel is designed for RWKV6, HEAD_SIZE == 64
-
-    dpct::queue_ptr stream = ctx.stream();
-
-    // Calculate execution configuration
-    const size_t shared_mem_size = WKV_BLOCK_SIZE * 4 * sizeof(float); // For k, r, tf, td
-    sycl::range<3> block_dims(1, 1, C / H);
-    sycl::range<3> grid_dims(1, 1, B * H);
-
-    // Submit kernel
-    stream->submit([&](sycl::handler& cgh) {
-        sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
-
-        cgh.parallel_for(
-            sycl::nd_range<3>(grid_dims * block_dims, block_dims),
-            [=](sycl::nd_item<3> item_ct1) {
-                rwkv_wkv_f32_kernel(
-                    B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
-                    item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
-                );
-            });
-    });
-
-    GGML_UNUSED(src0);
-    GGML_UNUSED(src1);
-}

package/src/llama.cpp/ggml/src/ggml-sycl/wkv6.hpp

@@ -1,9 +0,0 @@
-#ifndef GGML_SYCL_WKV6_HPP
-#define GGML_SYCL_WKV6_HPP
-
-#include "common.hpp"
-
-void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context & ctx, ggml_tensor * dst);
-
-
-#endif // GGML_SYCL_WKV6_HPP