Compare commits

..

1 Commits

Author SHA1 Message Date
Georgi Gerganov
2a615b27e4 ggml : remove redundant src in ggml_cast 2025-12-09 11:16:15 +02:00
5 changed files with 101 additions and 745 deletions

View File

@@ -1,11 +1,6 @@
#include "console.h"
#include <vector>
#include <iostream>
#include <cassert>
#include <cstddef>
#include <cctype>
#include <cwctype>
#include <cstdint>
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
@@ -40,26 +35,9 @@
namespace console {
#if defined (_WIN32)
namespace {
// Use private-use unicode values to represent special keys that are not reported
// as characters (e.g. arrows on Windows). These values should never clash with
// real input and let the rest of the code handle navigation uniformly.
static constexpr char32_t KEY_ARROW_LEFT = 0xE000;
static constexpr char32_t KEY_ARROW_RIGHT = 0xE001;
static constexpr char32_t KEY_ARROW_UP = 0xE002;
static constexpr char32_t KEY_ARROW_DOWN = 0xE003;
static constexpr char32_t KEY_HOME = 0xE004;
static constexpr char32_t KEY_END = 0xE005;
static constexpr char32_t KEY_CTRL_ARROW_LEFT = 0xE006;
static constexpr char32_t KEY_CTRL_ARROW_RIGHT = 0xE007;
static constexpr char32_t KEY_DELETE = 0xE008;
}
//
// Console state
//
#endif
static bool advanced_display = false;
static bool simple_io = true;
@@ -198,18 +176,7 @@ namespace console {
if (record.EventType == KEY_EVENT && record.Event.KeyEvent.bKeyDown) {
wchar_t wc = record.Event.KeyEvent.uChar.UnicodeChar;
if (wc == 0) {
const DWORD ctrl_mask = LEFT_CTRL_PRESSED | RIGHT_CTRL_PRESSED;
const bool ctrl_pressed = (record.Event.KeyEvent.dwControlKeyState & ctrl_mask) != 0;
switch (record.Event.KeyEvent.wVirtualKeyCode) {
case VK_LEFT: return ctrl_pressed ? KEY_CTRL_ARROW_LEFT : KEY_ARROW_LEFT;
case VK_RIGHT: return ctrl_pressed ? KEY_CTRL_ARROW_RIGHT : KEY_ARROW_RIGHT;
case VK_UP: return KEY_ARROW_UP;
case VK_DOWN: return KEY_ARROW_DOWN;
case VK_HOME: return KEY_HOME;
case VK_END: return KEY_END;
case VK_DELETE: return KEY_DELETE;
default: continue;
}
continue;
}
if ((wc >= 0xD800) && (wc <= 0xDBFF)) { // Check if wc is a high surrogate
@@ -348,52 +315,6 @@ namespace console {
#endif
}
static char32_t decode_utf8(const std::string & input, size_t pos, size_t & advance) {
unsigned char c = static_cast<unsigned char>(input[pos]);
if ((c & 0x80u) == 0u) {
advance = 1;
return c;
}
if ((c & 0xE0u) == 0xC0u && pos + 1 < input.size()) {
unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
if ((c1 & 0xC0u) != 0x80u) {
advance = 1;
return 0xFFFD;
}
advance = 2;
return ((c & 0x1Fu) << 6) | (static_cast<unsigned char>(input[pos + 1]) & 0x3Fu);
}
if ((c & 0xF0u) == 0xE0u && pos + 2 < input.size()) {
unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u) {
advance = 1;
return 0xFFFD;
}
advance = 3;
return ((c & 0x0Fu) << 12) |
((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 6) |
(static_cast<unsigned char>(input[pos + 2]) & 0x3Fu);
}
if ((c & 0xF8u) == 0xF0u && pos + 3 < input.size()) {
unsigned char c1 = static_cast<unsigned char>(input[pos + 1]);
unsigned char c2 = static_cast<unsigned char>(input[pos + 2]);
unsigned char c3 = static_cast<unsigned char>(input[pos + 3]);
if ((c1 & 0xC0u) != 0x80u || (c2 & 0xC0u) != 0x80u || (c3 & 0xC0u) != 0x80u) {
advance = 1;
return 0xFFFD;
}
advance = 4;
return ((c & 0x07u) << 18) |
((static_cast<unsigned char>(input[pos + 1]) & 0x3Fu) << 12) |
((static_cast<unsigned char>(input[pos + 2]) & 0x3Fu) << 6) |
(static_cast<unsigned char>(input[pos + 3]) & 0x3Fu);
}
advance = 1;
return 0xFFFD; // replacement character for invalid input
}
static void append_utf8(char32_t ch, std::string & out) {
if (ch <= 0x7F) {
out.push_back(static_cast<unsigned char>(ch));
@@ -415,319 +336,22 @@ namespace console {
}
// Helper function to remove the last UTF-8 character from a string
static size_t prev_utf8_char_pos(const std::string & line, size_t pos) {
if (pos == 0) return 0;
pos--;
while (pos > 0 && (line[pos] & 0xC0) == 0x80) {
pos--;
}
return pos;
}
static size_t next_utf8_char_pos(const std::string & line, size_t pos) {
if (pos >= line.length()) return line.length();
pos++;
while (pos < line.length() && (line[pos] & 0xC0) == 0x80) {
pos++;
}
return pos;
}
static void move_cursor(int delta);
static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths);
static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line);
static void delete_at_cursor(std::string & line, std::vector<int> & widths, size_t & char_pos, size_t & byte_pos) {
if (char_pos >= widths.size()) {
static void pop_back_utf8_char(std::string & line) {
if (line.empty()) {
return;
}
size_t next_pos = next_utf8_char_pos(line, byte_pos);
int w = widths[char_pos];
size_t char_len = next_pos - byte_pos;
size_t pos = line.length() - 1;
line.erase(byte_pos, char_len);
widths.erase(widths.begin() + char_pos);
size_t p = byte_pos;
int tail_width = 0;
for (size_t i = char_pos; i < widths.size(); ++i) {
size_t following = next_utf8_char_pos(line, p);
put_codepoint(line.c_str() + p, following - p, widths[i]);
tail_width += widths[i];
p = following;
// Find the start of the last UTF-8 character (checking up to 4 bytes back)
for (size_t i = 0; i < 3 && pos > 0; ++i, --pos) {
if ((line[pos] & 0xC0) != 0x80) {
break; // Found the start of the character
}
}
for (int i = 0; i < w; ++i) {
fputc(' ', out);
}
move_cursor(-(tail_width + w));
line.erase(pos);
}
static void clear_current_line(const std::vector<int> & widths) {
int total_width = 0;
for (int w : widths) {
total_width += (w > 0 ? w : 1);
}
if (total_width > 0) {
std::string spaces(total_width, ' ');
fwrite(spaces.c_str(), 1, total_width, out);
move_cursor(-total_width);
}
}
static void set_line_contents(std::string new_line, std::string & line, std::vector<int> & widths, size_t & char_pos,
size_t & byte_pos) {
move_to_line_start(char_pos, byte_pos, widths);
clear_current_line(widths);
line = std::move(new_line);
widths.clear();
byte_pos = 0;
char_pos = 0;
size_t idx = 0;
while (idx < line.size()) {
size_t advance = 0;
char32_t cp = decode_utf8(line, idx, advance);
int expected_width = estimateWidth(cp);
int real_width = put_codepoint(line.c_str() + idx, advance, expected_width);
if (real_width < 0) real_width = 0;
widths.push_back(real_width);
idx += advance;
++char_pos;
byte_pos = idx;
}
}
static void move_to_line_start(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths) {
int back_width = 0;
for (size_t i = 0; i < char_pos; ++i) {
back_width += widths[i];
}
move_cursor(-back_width);
char_pos = 0;
byte_pos = 0;
}
static void move_to_line_end(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
int forward_width = 0;
for (size_t i = char_pos; i < widths.size(); ++i) {
forward_width += widths[i];
}
move_cursor(forward_width);
char_pos = widths.size();
byte_pos = line.length();
}
static bool has_ctrl_modifier(const std::string & params) {
size_t start = 0;
while (start < params.size()) {
size_t end = params.find(';', start);
size_t len = (end == std::string::npos) ? params.size() - start : end - start;
if (len > 0) {
int value = 0;
for (size_t i = 0; i < len; ++i) {
char ch = params[start + i];
if (!std::isdigit(static_cast<unsigned char>(ch))) {
value = -1;
break;
}
value = value * 10 + (ch - '0');
}
if (value == 5) {
return true;
}
}
if (end == std::string::npos) {
break;
}
start = end + 1;
}
return false;
}
static bool is_space_codepoint(char32_t cp) {
return std::iswspace(static_cast<wint_t>(cp)) != 0;
}
static void move_word_left(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
if (char_pos == 0) {
return;
}
size_t new_char_pos = char_pos;
size_t new_byte_pos = byte_pos;
int move_width = 0;
while (new_char_pos > 0) {
size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
size_t advance = 0;
char32_t cp = decode_utf8(line, prev_byte, advance);
if (!is_space_codepoint(cp)) {
break;
}
move_width += widths[new_char_pos - 1];
new_char_pos--;
new_byte_pos = prev_byte;
}
while (new_char_pos > 0) {
size_t prev_byte = prev_utf8_char_pos(line, new_byte_pos);
size_t advance = 0;
char32_t cp = decode_utf8(line, prev_byte, advance);
if (is_space_codepoint(cp)) {
break;
}
move_width += widths[new_char_pos - 1];
new_char_pos--;
new_byte_pos = prev_byte;
}
move_cursor(-move_width);
char_pos = new_char_pos;
byte_pos = new_byte_pos;
}
static void move_word_right(size_t & char_pos, size_t & byte_pos, const std::vector<int> & widths, const std::string & line) {
if (char_pos >= widths.size()) {
return;
}
size_t new_char_pos = char_pos;
size_t new_byte_pos = byte_pos;
int move_width = 0;
while (new_char_pos < widths.size()) {
size_t advance = 0;
char32_t cp = decode_utf8(line, new_byte_pos, advance);
if (!is_space_codepoint(cp)) {
break;
}
move_width += widths[new_char_pos];
new_char_pos++;
new_byte_pos += advance;
}
while (new_char_pos < widths.size()) {
size_t advance = 0;
char32_t cp = decode_utf8(line, new_byte_pos, advance);
if (is_space_codepoint(cp)) {
break;
}
move_width += widths[new_char_pos];
new_char_pos++;
new_byte_pos += advance;
}
while (new_char_pos < widths.size()) {
size_t advance = 0;
char32_t cp = decode_utf8(line, new_byte_pos, advance);
if (!is_space_codepoint(cp)) {
break;
}
move_width += widths[new_char_pos];
new_char_pos++;
new_byte_pos += advance;
}
move_cursor(move_width);
char_pos = new_char_pos;
byte_pos = new_byte_pos;
}
static void move_cursor(int delta) {
if (delta == 0) return;
#if defined(_WIN32)
if (hConsole != NULL) {
CONSOLE_SCREEN_BUFFER_INFO bufferInfo;
GetConsoleScreenBufferInfo(hConsole, &bufferInfo);
COORD newCursorPosition = bufferInfo.dwCursorPosition;
int width = bufferInfo.dwSize.X;
int newX = newCursorPosition.X + delta;
int newY = newCursorPosition.Y;
while (newX >= width) {
newX -= width;
newY++;
}
while (newX < 0) {
newX += width;
newY--;
}
newCursorPosition.X = newX;
newCursorPosition.Y = newY;
SetConsoleCursorPosition(hConsole, newCursorPosition);
}
#else
if (delta < 0) {
for (int i = 0; i < -delta; i++) fprintf(out, "\b");
} else {
for (int i = 0; i < delta; i++) fprintf(out, "\033[C");
}
#endif
}
struct history_t {
std::vector<std::string> entries;
size_t viewing_idx = SIZE_MAX;
std::string backup_line; // current line before viewing history
void add(const std::string & line) {
if (line.empty()) {
return;
}
// avoid duplicates with the last entry
if (entries.empty() || entries.back() != line) {
entries.push_back(line);
}
// also clear viewing state
end_viewing();
}
bool prev(std::string & cur_line) {
if (entries.empty()) {
return false;
}
if (viewing_idx == SIZE_MAX) {
return false;
}
if (viewing_idx > 0) {
viewing_idx--;
}
cur_line = entries[viewing_idx];
return true;
}
bool next(std::string & cur_line) {
if (entries.empty() || viewing_idx == SIZE_MAX) {
return false;
}
viewing_idx++;
if (viewing_idx >= entries.size()) {
cur_line = backup_line;
end_viewing();
} else {
cur_line = entries[viewing_idx];
}
return true;
}
void begin_viewing(const std::string & line) {
backup_line = line;
viewing_idx = entries.size();
}
void end_viewing() {
viewing_idx = SIZE_MAX;
backup_line.clear();
}
bool is_viewing() const {
return viewing_idx != SIZE_MAX;
}
} history;
static bool readline_advanced(std::string & line, bool multiline_input) {
if (out != stdout) {
fflush(stdout);
@@ -738,33 +362,8 @@ namespace console {
bool is_special_char = false;
bool end_of_stream = false;
size_t byte_pos = 0; // current byte index
size_t char_pos = 0; // current character index (one char can be multiple bytes)
char32_t input_char;
while (true) {
assert(char_pos <= byte_pos);
assert(char_pos <= widths.size());
auto history_prev = [&]() {
if (!history.is_viewing()) {
history.begin_viewing(line);
}
std::string new_line;
if (!history.prev(new_line)) {
return;
}
set_line_contents(new_line, line, widths, char_pos, byte_pos);
};
auto history_next = [&]() {
if (history.is_viewing()) {
std::string new_line;
if (!history.next(new_line)) {
return;
}
set_line_contents(new_line, line, widths, char_pos, byte_pos);
}
};
fflush(out); // Ensure all output is displayed before waiting for input
input_char = getchar32();
@@ -772,7 +371,7 @@ namespace console {
break;
}
if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D */) {
if (input_char == (char32_t) WEOF || input_char == 0x04 /* Ctrl+D*/) {
end_of_stream = true;
break;
}
@@ -785,71 +384,7 @@ namespace console {
if (input_char == '\033') { // Escape sequence
char32_t code = getchar32();
if (code == '[') {
std::string params;
while (true) {
code = getchar32();
if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~' || code == (char32_t) WEOF) {
break;
}
params.push_back(static_cast<char>(code));
}
const bool ctrl_modifier = has_ctrl_modifier(params);
if (code == 'D') { // left
if (ctrl_modifier) {
move_word_left(char_pos, byte_pos, widths, line);
} else if (char_pos > 0) {
int w = widths[char_pos - 1];
move_cursor(-w);
char_pos--;
byte_pos = prev_utf8_char_pos(line, byte_pos);
}
} else if (code == 'C') { // right
if (ctrl_modifier) {
move_word_right(char_pos, byte_pos, widths, line);
} else if (char_pos < widths.size()) {
int w = widths[char_pos];
move_cursor(w);
char_pos++;
byte_pos = next_utf8_char_pos(line, byte_pos);
}
} else if (code == 'H') { // home
move_to_line_start(char_pos, byte_pos, widths);
} else if (code == 'F') { // end
move_to_line_end(char_pos, byte_pos, widths, line);
} else if (code == 'A' || code == 'B') {
// up/down
if (code == 'A') {
history_prev();
is_special_char = false;
} else if (code == 'B') {
history_next();
is_special_char = false;
}
} else if ((code == '~' || (code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z')) && !params.empty()) {
std::string digits;
for (char ch : params) {
if (ch == ';') {
break;
}
if (std::isdigit(static_cast<unsigned char>(ch))) {
digits.push_back(ch);
}
}
if (code == '~') {
if (digits == "1" || digits == "7") { // home
move_to_line_start(char_pos, byte_pos, widths);
} else if (digits == "4" || digits == "8") { // end
move_to_line_end(char_pos, byte_pos, widths, line);
} else if (digits == "3") { // delete
delete_at_cursor(line, widths, char_pos, byte_pos);
}
}
}
} else if (code == 0x1B) {
if (code == '[' || code == 0x1B) {
// Discard the rest of the escape sequence
while ((code = getchar32()) != (char32_t) WEOF) {
if ((code >= 'A' && code <= 'Z') || (code >= 'a' && code <= 'z') || code == '~') {
@@ -857,107 +392,28 @@ namespace console {
}
}
}
#if defined(_WIN32)
} else if (input_char == KEY_ARROW_LEFT) {
if (char_pos > 0) {
int w = widths[char_pos - 1];
move_cursor(-w);
char_pos--;
byte_pos = prev_utf8_char_pos(line, byte_pos);
}
} else if (input_char == KEY_ARROW_RIGHT) {
if (char_pos < widths.size()) {
int w = widths[char_pos];
move_cursor(w);
char_pos++;
byte_pos = next_utf8_char_pos(line, byte_pos);
}
} else if (input_char == KEY_CTRL_ARROW_LEFT) {
move_word_left(char_pos, byte_pos, widths, line);
} else if (input_char == KEY_CTRL_ARROW_RIGHT) {
move_word_right(char_pos, byte_pos, widths, line);
} else if (input_char == KEY_HOME) {
move_to_line_start(char_pos, byte_pos, widths);
} else if (input_char == KEY_END) {
move_to_line_end(char_pos, byte_pos, widths, line);
} else if (input_char == KEY_DELETE) {
delete_at_cursor(line, widths, char_pos, byte_pos);
} else if (input_char == KEY_ARROW_UP || input_char == KEY_ARROW_DOWN) {
if (input_char == KEY_ARROW_UP) {
history_prev();
is_special_char = false;
} else if (input_char == KEY_ARROW_DOWN) {
history_next();
is_special_char = false;
}
#endif
} else if (input_char == 0x08 || input_char == 0x7F) { // Backspace
if (char_pos > 0) {
int w = widths[char_pos - 1];
move_cursor(-w);
char_pos--;
size_t prev_pos = prev_utf8_char_pos(line, byte_pos);
size_t char_len = byte_pos - prev_pos;
byte_pos = prev_pos;
// remove the character
line.erase(byte_pos, char_len);
widths.erase(widths.begin() + char_pos);
// redraw tail
size_t p = byte_pos;
int tail_width = 0;
for (size_t i = char_pos; i < widths.size(); ++i) {
size_t next_p = next_utf8_char_pos(line, p);
put_codepoint(line.c_str() + p, next_p - p, widths[i]);
tail_width += widths[i];
p = next_p;
}
// clear display
for (int i = 0; i < w; ++i) {
fputc(' ', out);
}
move_cursor(-(tail_width + w));
if (!widths.empty()) {
int count;
do {
count = widths.back();
widths.pop_back();
// Move cursor back, print space, and move cursor back again
for (int i = 0; i < count; i++) {
replace_last(' ');
pop_cursor();
}
pop_back_utf8_char(line);
} while (count == 0 && !widths.empty());
}
} else {
// insert character
std::string new_char_str;
append_utf8(input_char, new_char_str);
int w = estimateWidth(input_char);
if (char_pos == widths.size()) {
// insert at the end
line += new_char_str;
int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
if (real_w < 0) real_w = 0;
widths.push_back(real_w);
byte_pos += new_char_str.length();
char_pos++;
} else {
// insert in middle
line.insert(byte_pos, new_char_str);
int real_w = put_codepoint(new_char_str.c_str(), new_char_str.length(), w);
if (real_w < 0) real_w = 0;
widths.insert(widths.begin() + char_pos, real_w);
// print the tail
size_t p = byte_pos + new_char_str.length();
int tail_width = 0;
for (size_t i = char_pos + 1; i < widths.size(); ++i) {
size_t next_p = next_utf8_char_pos(line, p);
put_codepoint(line.c_str() + p, next_p - p, widths[i]);
tail_width += widths[i];
p = next_p;
}
move_cursor(-tail_width);
byte_pos += new_char_str.length();
char_pos++;
int offset = line.length();
append_utf8(input_char, line);
int width = put_codepoint(line.c_str() + offset, line.length() - offset, estimateWidth(input_char));
if (width < 0) {
width = 0;
}
widths.push_back(width);
}
if (!line.empty() && (line.back() == '\\' || line.back() == '/')) {
@@ -995,15 +451,6 @@ namespace console {
}
}
if (!end_of_stream && !line.empty()) {
// remove the trailing newline for history storage
if (!line.empty() && line.back() == '\n') {
line.pop_back();
}
// TODO: maybe support multiline history entries?
history.add(line);
}
fflush(out);
return has_more;
}

View File

@@ -2251,12 +2251,12 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
int sections[4],
bool mrope_used,
bool is_imrope,
bool indep_sects,
int64_t rope_dims) {
bool indep_sects) {
ggml_tensor * src0 = dst->src[0]; // input
ggml_tensor * src1 = dst->src[1]; // position
ggml_tensor * src2 = dst->src[2]; // freq_factors
int64_t theta_scale_length = rope_dims / 2;
int64_t theta_scale_length = src0->ne[0] / 2;
int64_t position_length = dst->ne[2];
// TODO: check theta_scale_length and position_length.
@@ -2331,17 +2331,18 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
ACL_CHECK(aclrtMemcpyAsync(ctx.rope_cache.theta_scale_cache, theta_scale_length * sizeof(float),
ctx.rope_cache.theta_scale_exp_host, theta_scale_length * sizeof(float),
ACL_MEMCPY_HOST_TO_DEVICE, ctx.stream()));
acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
theta_scale_ne, theta_scale_nb, 1);
}
acl_theta_scale_tensor = ggml_cann_create_tensor(ctx.rope_cache.theta_scale_cache, ACL_FLOAT, sizeof(float),
theta_scale_ne, theta_scale_nb, 1);
// Step1.2: prepare rope_yarn_ramp, if this part updated, should update theta_scale_tensor.
// TODO: acl_yarn_ramp_tensor use rope cache.
bool yarn_ramp_tensor_updated = false;
ggml_cann_pool_alloc yarn_ramp_allocator(ctx.pool());
acl_tensor_ptr acl_yarn_ramp_tensor;
if (ext_factor != 0 && (theta_scale_updated || ctx.rope_cache.theta_scale_length != theta_scale_length ||
ctx.rope_cache.freq_scale != freq_scale)) {
if (ext_factor != 0 &&
// TODO: check more parameter.
(ctx.rope_cache.theta_scale_length != theta_scale_length || ctx.rope_cache.freq_scale != freq_scale)) {
yarn_ramp_tensor_updated = true;
// -rope_yarn_ramp
@@ -2589,7 +2590,7 @@ static void aclnn_rope_cache_init(ggml_backend_cann_context & ctx,
aclnn_muls(ctx, acl_cos_tensor.get(), attn_factor, nullptr, true);
}
int64_t sin_reshape_ne[4] = { rope_dims, 1, dst->ne[2], 1 };
int64_t sin_reshape_ne[4] = { src0->ne[0], 1, dst->ne[2], 1 };
size_t sin_reshape_nb[GGML_MAX_DIMS];
sin_reshape_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2644,7 +2645,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
// param
float freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow;
int sections[4];
int sections[4];
// const int n_past = ((int32_t *) dst->op_params)[0];
const int n_dims = ((int32_t *) dst->op_params)[1];
const int mode = ((int32_t *) dst->op_params)[2];
@@ -2653,60 +2654,44 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
GGML_TENSOR_UNARY_OP_LOCALS
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
memcpy(&sections, (int32_t *) dst->op_params + 11, sizeof(int) * 4);
memcpy(&freq_base, (int32_t *) dst->op_params + 5, sizeof(float));
memcpy(&freq_scale, (int32_t *) dst->op_params + 6, sizeof(float));
memcpy(&ext_factor, (int32_t *) dst->op_params + 7, sizeof(float));
memcpy(&attn_factor, (int32_t *) dst->op_params + 8, sizeof(float));
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
memcpy(&sections, (int32_t *) dst->op_params + 11, sizeof(int)*4);
// TODO: n_dims <= ne0
GGML_ASSERT(n_dims == ne0);
GGML_ASSERT(n_dims % 2 == 0);
GGML_ASSERT(n_dims <= ne00);
const float theta_scale = powf(freq_base, -2.0f / n_dims);
float corr_dims[2];
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast, beta_slow, corr_dims);
bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
// mrope_used means the GGML_ROPE_TYPE_MROPE bit is set.
// Note: this bit is also set for imrope and some vision modes,
// so mrope_used does NOT exclusively indicate pure mrope.
const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE;
const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const bool is_imrope = mode == GGML_ROPE_TYPE_IMROPE; // qwen3vl apply interleaved mrope
const bool mrope_used = mode & GGML_ROPE_TYPE_MROPE; // ggml_rope_multi, note: also true for vision (24 & 8 == true) and for imrope
const bool is_vision = mode == GGML_ROPE_TYPE_VISION;
if (mrope_used) {
GGML_ASSERT(sections[0] > 0 || sections[1] > 0 || sections[2] > 0);
}
if (is_vision) {
GGML_ASSERT(n_dims == ne0 / 2);
GGML_ASSERT(n_dims == ne0/2);
}
if (is_imrope || mrope_used) {
is_neox = true;
}
int64_t rope_dims = n_dims;
//Our current RotaryPositionEmbedding does not support the VISION mode,
//but essentially it only modifies theta_base in mrope,
//then repeats it at the end in the same way as is_neox.
//In fact, RoPE is still applied across all dimensions.
if (is_vision) {
rope_dims = src0->ne[0];
}
int64_t tail_dims = ne00 - rope_dims;
bool has_tail = tail_dims > 0;
// init ctx.rope_cos/rope_sin cache
aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections,
mrope_used, is_imrope, is_vision, rope_dims);
aclnn_rope_cache_init(ctx, dst, corr_dims, ext_factor, theta_scale, freq_scale, attn_factor, is_neox, sections, mrope_used, is_imrope, is_vision);
// Cache is generated with ne00 dimensions, so we use ne00 for reshape
int64_t sin_reshape_ne[4] = { rope_dims, 1, ne02, 1 };
int64_t sin_reshape_ne[4] = { ne00, 1, ne02, 1 };
size_t sin_reshape_nb[GGML_MAX_DIMS];
sin_reshape_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
@@ -2719,6 +2704,7 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
acl_tensor_ptr acl_src = ggml_cann_create_tensor(src0);
acl_tensor_ptr acl_dst = ggml_cann_create_tensor(dst);
#ifdef ASCEND_310P
// Special ROPE operation for 310P
@@ -2858,124 +2844,46 @@ void ggml_cann_rope(ggml_backend_cann_context & ctx, ggml_tensor * dst) {
}
return;
#endif
int64_t acl_mode = is_neox ? 0 : 1;
// Pre-define head and tail dimensions for reuse
int64_t head_ne[GGML_MAX_DIMS] = { rope_dims, ne01, ne02, ne03 };
int64_t tail_ne[GGML_MAX_DIMS] = { tail_dims, ne01, ne02, ne03 };
switch (src0->type) {
case GGML_TYPE_F32:
{
GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src.get(), acl_cos_reshape_tensor.get(),
acl_sin_reshape_tensor.get(), acl_mode, acl_dst.get());
break;
}
case GGML_TYPE_F16:
{
ggml_cann_pool_alloc src_trans_allocator(ctx.pool(), ggml_nelements(src0) * sizeof(float));
void * src_trans_buffer = src_trans_allocator.get();
ggml_cann_pool_alloc dst_trans_allocator(ctx.pool(), ggml_nelements(dst) * sizeof(float));
void * dst_trans_buffer = dst_trans_allocator.get();
// Step 1: Prepare trans tensors for F16 type conversion to F32 if needed
bool src_dst_need_trans = false;
ggml_cann_pool_alloc src_trans_allocator(ctx.pool());
ggml_cann_pool_alloc dst_trans_allocator(ctx.pool());
acl_tensor_ptr acl_src_trans_tensor;
acl_tensor_ptr acl_dst_trans_tensor;
void * src_trans_buffer = nullptr;
void * dst_trans_buffer = nullptr;
size_t src_dst_trans_nb[GGML_MAX_DIMS];
if (src0->type == GGML_TYPE_F16) {
src_dst_need_trans = true;
src_trans_buffer = src_trans_allocator.alloc(ggml_nelements(src0) * sizeof(float));
dst_trans_buffer = dst_trans_allocator.alloc(ggml_nelements(dst) * sizeof(float));
size_t src_trans_nb[GGML_MAX_DIMS];
src_trans_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_trans_nb[i] = src_trans_nb[i - 1] * src0->ne[i - 1];
}
src_dst_trans_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
src_dst_trans_nb[i] = src_dst_trans_nb[i - 1] * src0->ne[i - 1];
}
acl_src_trans_tensor = ggml_cann_create_tensor(src_trans_buffer, ACL_FLOAT, sizeof(float), src0->ne,
src_dst_trans_nb, GGML_MAX_DIMS);
acl_dst_trans_tensor = ggml_cann_create_tensor(dst_trans_buffer, ACL_FLOAT, sizeof(float), dst->ne,
src_dst_trans_nb, GGML_MAX_DIMS);
aclnn_cast(ctx, acl_src.get(), acl_src_trans_tensor.get(), ACL_FLOAT);
}
acl_tensor_ptr acl_src_trans_tensor = ggml_cann_create_tensor(
src_trans_buffer, ACL_FLOAT, sizeof(float), src0->ne, src_trans_nb, GGML_MAX_DIMS);
acl_tensor_ptr acl_dst_trans_tensor = ggml_cann_create_tensor(
dst_trans_buffer, ACL_FLOAT, sizeof(float), dst->ne, src_trans_nb, GGML_MAX_DIMS);
// Step 2: Prepare head tensors for tail splitting if needed
acl_tensor_ptr acl_src_head;
acl_tensor_ptr acl_dst_head;
if (has_tail) {
// Create head views for RotaryPositionEmbedding (only first rope_dims dimensions)
// RotaryPositionEmbedding requires contiguous dst tensor, so we use a temporary buffer
if (src_dst_need_trans) {
// Use F32 trans tensor strides
acl_src_head = ggml_cann_create_tensor((char *) src_trans_buffer, ACL_FLOAT, sizeof(float), head_ne,
src_dst_trans_nb, GGML_MAX_DIMS);
} else {
// Use original F32 tensor strides
acl_src_head = ggml_cann_create_tensor((char *) src0->data, ACL_FLOAT, sizeof(float), head_ne, src0->nb,
GGML_MAX_DIMS);
}
aclnn_cast(ctx, acl_src.get(), acl_src_trans_tensor.get(), ACL_FLOAT);
int64_t head_elements = rope_dims * ne01 * ne02 * ne03;
ggml_cann_pool_alloc dst_head_contiguous_allocator(ctx.pool(), head_elements * sizeof(float));
void * dst_head_contiguous_buffer = dst_head_contiguous_allocator.get();
GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_trans_tensor.get(),
acl_cos_reshape_tensor.get(), acl_sin_reshape_tensor.get(), acl_mode,
acl_dst_trans_tensor.get());
size_t head_contiguous_nb[GGML_MAX_DIMS];
head_contiguous_nb[0] = sizeof(float);
for (int i = 1; i < GGML_MAX_DIMS; i++) {
head_contiguous_nb[i] = head_contiguous_nb[i - 1] * head_ne[i - 1];
}
acl_dst_head = ggml_cann_create_tensor(dst_head_contiguous_buffer, ACL_FLOAT, sizeof(float), head_ne,
head_contiguous_nb, GGML_MAX_DIMS);
}
// Step 3: Execute RotaryPositionEmbedding
if (has_tail) {
// Rotate only the head portion (first rope_dims dimensions)
GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_head.get(), acl_cos_reshape_tensor.get(),
acl_sin_reshape_tensor.get(), acl_mode, acl_dst_head.get());
// Copy head result from contiguous buffer back to destination tensor
if (src_dst_need_trans) {
acl_tensor_ptr acl_dst_head_target = ggml_cann_create_tensor(
(char *) dst_trans_buffer, ACL_FLOAT, sizeof(float), head_ne, src_dst_trans_nb, GGML_MAX_DIMS);
cann_copy(ctx, acl_dst_head.get(), acl_dst_head_target.get());
} else {
acl_tensor_ptr acl_dst_head_target =
ggml_cann_create_tensor((char *) dst->data, ACL_FLOAT, sizeof(float), head_ne, dst->nb, GGML_MAX_DIMS);
cann_copy(ctx, acl_dst_head.get(), acl_dst_head_target.get());
}
} else if (src_dst_need_trans) {
// Rotate full tensor (no tail), using trans tensors
GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src_trans_tensor.get(), acl_cos_reshape_tensor.get(),
acl_sin_reshape_tensor.get(), acl_mode, acl_dst_trans_tensor.get());
} else {
// Rotate full tensor (no tail), using original tensors
GGML_CANN_CALL_ACLNN_OP(ctx, RotaryPositionEmbedding, acl_src.get(), acl_cos_reshape_tensor.get(),
acl_sin_reshape_tensor.get(), acl_mode, acl_dst.get());
}
// Step 4: Copy unrotated tail portion from source to destination
if (has_tail) {
size_t src_tail_offset;
size_t dst_tail_offset;
auto copy_tail_device = [&](void * src_ptr, void * dst_ptr, aclDataType dtype, size_t elem_size,
size_t * nb_src_arr, size_t * nb_dst_arr) {
acl_tensor_ptr acl_src_tail =
ggml_cann_create_tensor(src_ptr, dtype, elem_size, tail_ne, nb_src_arr, GGML_MAX_DIMS);
acl_tensor_ptr acl_dst_tail =
ggml_cann_create_tensor(dst_ptr, dtype, elem_size, tail_ne, nb_dst_arr, GGML_MAX_DIMS);
cann_copy(ctx, acl_src_tail.get(), acl_dst_tail.get());
};
if (src_dst_need_trans) {
// Use F32 trans tensor strides and offsets
src_tail_offset = rope_dims * src_dst_trans_nb[0];
dst_tail_offset = rope_dims * src_dst_trans_nb[0];
copy_tail_device((char *) src_trans_buffer + src_tail_offset, (char *) dst_trans_buffer + dst_tail_offset,
ACL_FLOAT, sizeof(float), src_dst_trans_nb, src_dst_trans_nb);
} else {
// Use original tensor strides and offsets
src_tail_offset = rope_dims * nb00;
dst_tail_offset = rope_dims * nb0;
copy_tail_device((char *) src0->data + src_tail_offset, (char *) dst->data + dst_tail_offset,
ggml_cann_type_mapping(dst->type), ggml_element_size(dst), src0->nb, dst->nb);
}
}
// Step 5: Cast back to F16 if needed
if (src_dst_need_trans) {
aclnn_cast(ctx, acl_dst_trans_tensor.get(), acl_dst.get(), ACL_FLOAT16);
aclnn_cast(ctx, acl_dst_trans_tensor.get(), acl_dst.get(), ACL_FLOAT16);
break;
}
default:
GGML_ABORT("Unsupported tensor type for GGML_OP_ROPE");
break;
}
}

View File

@@ -315,7 +315,7 @@ struct ggml_cann_rope_cache {
if (theta_scale_exp_host) {
free(theta_scale_exp_host);
}
if (position_select_index_host) {
if(position_select_index_host) {
free(position_select_index_host);
}
}
@@ -340,7 +340,7 @@ struct ggml_cann_rope_cache {
void set(int64_t theta_scale_length,
int64_t position_length,
float ext_factor,
float ext_factor,
float theta_scale,
float freq_scale,
float attn_factor,

View File

@@ -2308,7 +2308,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
bool cann_graph_update_required = false;
#ifdef USE_ACL_GRAPH
bool use_cann_graph = true;
bool use_cann_graph = true;
static bool prefill_use_graph = parse_bool(get_env("GGML_CANN_PREFILL_USE_GRAPH").value_or(""));
if (!prefill_use_graph) {
@@ -2338,7 +2338,7 @@ static enum ggml_status ggml_backend_cann_graph_compute(ggml_backend_t backend,
}
}
#else
bool use_cann_graph = false;
bool use_cann_graph = false;
#endif // USE_ACL_GRAPH
evaluate_and_capture_cann_graph(cann_ctx, cgraph, use_cann_graph, cann_graph_update_required);
@@ -2474,14 +2474,16 @@ static bool ggml_backend_cann_supports_op(ggml_backend_dev_t dev, const ggml_ten
}
case GGML_OP_ROPE:
{
// TODO: with ops-test v == 1
// TODO: n_dims <= ne0
if (op->src[0]->ne[0] != op->op_params[1]) {
return false;
}
if (op->src[0]->ne[0] > 896) {
return false;
}
#ifdef ASCEND_310P
// TODO: Support rope_dim < ne00(dim)
if (op->src[0]->ne[0] != op->op_params[1]) {
return false;
}
if (!ggml_is_contiguous(op->src[0])) {
return false;
}

View File

@@ -3418,7 +3418,6 @@ struct ggml_tensor * ggml_cast(
result->op = GGML_OP_CPY;
result->src[0] = a;
result->src[1] = result;
return result;
}