add inplace conversion support for f8_e4m3 (#359)

in the same way it is done for bf16
like how bf16 converts losslessly to fp32,
f8_e4m3 converts losslessly to fp16

Author: Green-Sky

model.cpp

@@ -554,13 +554,62 @@ float bf16_to_f32(uint16_t bfloat16) {
     return *reinterpret_cast<float*>(&val_bits);
 }
 
+uint16_t f8_e4m3_to_f16(uint8_t f8) {
+    // do we need to support uz?
+
+    const uint32_t exponent_bias = 7;
+    if (f8 == 0xff) {
+        return ggml_fp32_to_fp16(-NAN);
+    } else if (f8 == 0x7f) {
+        return ggml_fp32_to_fp16(NAN);
+    }
+
+    uint32_t sign = f8 & 0x80;
+    uint32_t exponent = (f8 & 0x78) >> 3;
+    uint32_t mantissa = f8 & 0x07;
+    uint32_t result = sign << 24;
+    if (exponent == 0) {
+        if (mantissa > 0) {
+            exponent = 0x7f - exponent_bias;
+
+            // yes, 2 times
+            if ((mantissa & 0x04) == 0) {
+                mantissa &= 0x03;
+                mantissa <<= 1;
+                exponent -= 1;
+            }
+            if ((mantissa & 0x04) == 0) {
+                mantissa &= 0x03;
+                mantissa <<= 1;
+                exponent -= 1;
+            }
+
+            result |= (mantissa & 0x03) << 21;
+            result |= exponent << 23;
+        }
+    } else {
+        result |= mantissa << 20;
+        exponent += 0x7f - exponent_bias;
+        result |= exponent << 23;
+    }
+
+    return ggml_fp32_to_fp16(*reinterpret_cast<const float*>(&result));
+}
+
 void bf16_to_f32_vec(uint16_t* src, float* dst, int64_t n) {
     // support inplace op
     for (int64_t i = n - 1; i >= 0; i--) {
         dst[i] = bf16_to_f32(src[i]);
     }
 }
 
+void f8_e4m3_to_f16_vec(uint8_t* src, uint16_t* dst, int64_t n) {
+    // support inplace op
+    for (int64_t i = n - 1; i >= 0; i--) {
+        dst[i] = f8_e4m3_to_f16(src[i]);
+    }
+}
+
 void convert_tensor(void* src,
                     ggml_type src_type,
                     void* dst,
@@ -794,6 +843,8 @@ ggml_type str_to_ggml_type(const std::string& dtype) {
         ttype = GGML_TYPE_F32;
     } else if (dtype == "F32") {
         ttype = GGML_TYPE_F32;
+    } else if (dtype == "F8_E4M3") {
+        ttype = GGML_TYPE_F16;
     }
     return ttype;
 }
@@ -866,7 +917,7 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
 
         ggml_type type = str_to_ggml_type(dtype);
         if (type == GGML_TYPE_COUNT) {
-            LOG_ERROR("unsupported dtype '%s'", dtype.c_str());
+            LOG_ERROR("unsupported dtype '%s' (tensor '%s')", dtype.c_str(), name.c_str());
             return false;
         }
 
@@ -903,6 +954,10 @@ bool ModelLoader::init_from_safetensors_file(const std::string& file_path, const
         if (dtype == "BF16") {
             tensor_storage.is_bf16 = true;
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
+        } else if (dtype == "F8_E4M3") {
+            tensor_storage.is_f8_e4m3 = true;
+            // f8 -> f16
+            GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size * 2);
         } else {
             GGML_ASSERT(tensor_storage.nbytes() == tensor_data_size);
         }
@@ -1537,6 +1592,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     if (tensor_storage.is_bf16) {
                         // inplace op
                         bf16_to_f32_vec((uint16_t*)dst_tensor->data, (float*)dst_tensor->data, tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e4m3) {
+                        // inplace op
+                        f8_e4m3_to_f16_vec((uint8_t*)dst_tensor->data, (uint16_t*)dst_tensor->data, tensor_storage.nelements());
                     }
                 } else {
                     read_buffer.resize(tensor_storage.nbytes());
@@ -1545,6 +1603,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                     if (tensor_storage.is_bf16) {
                         // inplace op
                         bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                    } else if (tensor_storage.is_f8_e4m3) {
+                        // inplace op
+                        f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                     }
 
                     convert_tensor((void*)read_buffer.data(), tensor_storage.type, dst_tensor->data,
@@ -1557,6 +1618,9 @@ bool ModelLoader::load_tensors(on_new_tensor_cb_t on_new_tensor_cb, ggml_backend
                 if (tensor_storage.is_bf16) {
                     // inplace op
                     bf16_to_f32_vec((uint16_t*)read_buffer.data(), (float*)read_buffer.data(), tensor_storage.nelements());
+                } else if (tensor_storage.is_f8_e4m3) {
+                    // inplace op
+                    f8_e4m3_to_f16_vec((uint8_t*)read_buffer.data(), (uint16_t*)read_buffer.data(), tensor_storage.nelements());
                 }
 
                 if (tensor_storage.type == dst_tensor->type) {

model.h

@@ -32,6 +32,7 @@ struct TensorStorage {
     std::string name;
     ggml_type type          = GGML_TYPE_F32;
     bool is_bf16            = false;
+    bool is_f8_e4m3         = false;
     int64_t ne[SD_MAX_DIMS] = {1, 1, 1, 1, 1};
     int n_dims              = 0;
 
@@ -61,7 +62,7 @@ struct TensorStorage {
     }
 
     int64_t nbytes_to_read() const {
-        if (is_bf16) {
+        if (is_bf16 || is_f8_e4m3) {
             return nbytes() / 2;
         } else {
             return nbytes();
@@ -109,6 +110,8 @@ struct TensorStorage {
         const char* type_name = ggml_type_name(type);
         if (is_bf16) {
             type_name = "bf16";
+        } else if (is_f8_e4m3) {
+            type_name = "f8_e4m3";
         }
         ss << name << " | " << type_name << " | ";
         ss << n_dims << " [";
@@ -160,4 +163,6 @@ class ModelLoader {
     static std::string load_merges();
     static std::string load_t5_tokenizer_json();
 };
-#endif  // __MODEL_H__
+
+#endif  // __MODEL_H__
+