Merge remote-tracking branch 'origin/master' into tool-bench-prod

Author: Olivier Chafik

docs/function-calling.md

@@ -521,8 +521,13 @@ static json oaicompat_completion_params_parse(const json & body) {
         throw std::runtime_error("Only one completion choice is allowed");
     }
 
+    // Handle "echo" field
+    if (json_value(body, "echo", false)) {
+        throw std::runtime_error("Only no echo is supported");
+    }
+
     // Params supported by OAI but unsupported by llama.cpp
-    static const std::vector<std::string> unsupported_params { "best_of", "echo", "suffix" };
+    static const std::vector<std::string> unsupported_params { "best_of", "suffix" };
     for (const auto & param : unsupported_params) {
         if (body.contains(param)) {
             throw std::runtime_error("Unsupported param: " + param);

examples/server/README.md

@@ -5265,6 +5265,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
 
 #if defined(__ARM_FEATURE_SVE)
 
+    uint32_t aux[3];
     uint32_t utmp[4];
 
     const int8_t m32 = 32;
@@ -5276,7 +5277,6 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
     const svuint8_t m1_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 1);
     const svuint8_t m2_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 2);
     const svuint8_t m3_sv = svlsl_n_u8_x(svptrue_b8(), m0_sv, 3);
-    svbool_t pred_s32 = svnot_b_z (svptrue_b32(), svptrue_pat_b32(SV_VL4));
 
     float sum = 0;
 
@@ -5289,7 +5289,7 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, size_t bs, const void * r
         const int8_t  * restrict q8_sv = y[i].qs;
 
         // Set up scales
-        uint32_t * aux = &x[i].scales;
+        memcpy(aux, x[i].scales, 12);
         utmp[3] = ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4);
         utmp[2] = ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4);
         utmp[1] = (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4);

examples/server/utils.hpp

@@ -0,0 +1,55 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+#define BLOCK_SIZE 512
+
+layout(local_size_x = BLOCK_SIZE, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer G {A_TYPE data_a[];};
+layout (binding = 1) readonly buffer X {B_TYPE data_b[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE sum_xx[BLOCK_SIZE];
+shared FLOAT_TYPE sum_xg[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    // Compute derivative of x[i]/norm(x) = g[i]/norm(x) - x[i] dot(x,g)/KX / norm(x)^1.5
+
+    // partial sums for thread in warp
+    sum_xx[tid] = FLOAT_TYPE(0.0f);
+    sum_xg[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const FLOAT_TYPE gi = FLOAT_TYPE(data_a[row*p.KX + col]);
+        const FLOAT_TYPE xi = FLOAT_TYPE(data_b[row*p.KX + col]);
+        sum_xx[tid] += xi * xi;
+        sum_xg[tid] += xi * gi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (int s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum_xx[tid] += sum_xx[tid + s];
+            sum_xg[tid] += sum_xg[tid + s];
+        }
+        barrier();
+    }
+
+    const FLOAT_TYPE eps = FLOAT_TYPE(p.param1);
+    const FLOAT_TYPE mean = sum_xx[0] / FLOAT_TYPE(p.KX);
+    const FLOAT_TYPE scale_g = inversesqrt(mean + eps);
+    const FLOAT_TYPE scale_x = -scale_g * sum_xg[0] / (sum_xx[0] + FLOAT_TYPE(p.KX) * eps);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE(
+            scale_g * FLOAT_TYPE(data_a[row*p.KX + col]) +
+            scale_x * FLOAT_TYPE(data_b[row*p.KX + col]));
+    }
+}

ggml/src/ggml-cpu/ggml-cpu-quants.c

@@ -29,6 +29,7 @@ layout (push_constant) uniform parameter {
     uint s1;
     uint s2;
     int sections[4];
+    uint is_back;
 } p;
 
 float rope_yarn_ramp(const float low, const float high, const uint i0) {
@@ -48,6 +49,10 @@ void rope_yarn(const float theta_extrap, const uint i0, out float cos_theta, out
         // Get n-d magnitude scaling corrected for interpolation
         mscale *= 1.0f + 0.1f * log(1.0f / p.freq_scale);
     }
+    // Backprogagation uses inverted rotation
+    if (p.is_back != 0) {
+        theta = -theta;
+    }
     cos_theta = cos(theta) * mscale;
     sin_theta = sin(theta) * mscale;
 }

ggml/src/ggml-vulkan/ggml-vulkan.cpp

@@ -0,0 +1,20 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer X {A_TYPE data_a[];};
+layout (binding = 1) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+    data_d[i] = D_TYPE(1. / (1 + exp(-1. *data_a[i])));
+}

ggml/src/ggml-vulkan/vulkan-shaders/rms_norm_back.comp

@@ -0,0 +1,26 @@
+#version 450
+
+#include "generic_head.comp"
+#include "types.comp"
+
+#extension GL_EXT_control_flow_attributes : enable
+
+layout(local_size_x = 512, local_size_y = 1, local_size_z = 1) in;
+
+layout (binding = 0) readonly buffer G {A_TYPE data_g[];};
+layout (binding = 1) readonly buffer X {B_TYPE data_x[];};
+layout (binding = 2) writeonly buffer D {D_TYPE data_d[];};
+
+void main() {
+    const uint i = gl_GlobalInvocationID.z * 262144 + gl_GlobalInvocationID.y * 512 + gl_GlobalInvocationID.x;
+
+    if (i >= p.KX) {
+        return;
+    }
+
+    // Compute derivative of SiLU(x): 1/(1+exp(-x)) - x*exp(-x)/(1+exp(-x))^2
+
+    const float xi = float(data_x[i]);
+    const float s = 1.0f / (1.0f + exp(-xi));
+    data_d[i] = D_TYPE(data_g[i] * (s + xi * s * (1 - s)));
+}

ggml/src/ggml-vulkan/vulkan-shaders/rope_head.comp

@@ -0,0 +1,50 @@
+#version 450
+
+#extension GL_EXT_control_flow_attributes : enable
+
+#include "generic_head.comp"
+#include "types.comp"
+
+layout(constant_id = 0) const uint BLOCK_SIZE = 32;
+layout(local_size_x_id = 0, local_size_y = 1, local_size_z = 1) in;
+
+// In this shader Y = softmax(X) and X is not provided as input.
+
+layout (binding = 0) readonly buffer G {A_TYPE data_g[];};
+layout (binding = 1) readonly buffer Y {B_TYPE data_y[];};
+layout (binding = 2) buffer D {D_TYPE data_d[];};
+
+shared FLOAT_TYPE sum_yg[BLOCK_SIZE];
+
+void main() {
+    const uint row = gl_WorkGroupID.z * 262144 + gl_WorkGroupID.y * 512 + gl_WorkGroupID.x;
+    const uint tid = gl_LocalInvocationID.x;
+
+    FLOAT_TYPE scale = p.param1;
+
+    // partial sums for thread in warp
+    sum_yg[tid] = FLOAT_TYPE(0.0f);
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        const FLOAT_TYPE gi = FLOAT_TYPE(data_g[row*p.KX + col]);
+        const FLOAT_TYPE yi = FLOAT_TYPE(data_y[row*p.KX + col]);
+        sum_yg[tid] += yi * gi;
+    }
+
+    // sum up partial sums and write back result
+    barrier();
+    [[unroll]] for (uint s = BLOCK_SIZE / 2; s > 0; s >>= 1) {
+        if (tid < s) {
+            sum_yg[tid] += sum_yg[tid + s];
+        }
+        barrier();
+    }
+
+    const FLOAT_TYPE dot_yg = sum_yg[0];
+
+    [[unroll]] for (uint col = tid; col < p.KX; col += BLOCK_SIZE) {
+        data_d[row*p.KX + col] = D_TYPE(scale
+            * (FLOAT_TYPE(data_g[row*p.KX + col]) - dot_yg)
+            * FLOAT_TYPE(data_y[row*p.KX + col]));
+    }
+}