tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-17 13:40:55 -04:00
Code Issues Packages Projects Releases Wiki Activity

kv-cache : log (debug) all streams in find_slot (#15176)

Browse Source 
This commit updates `llama_kv_cache_unified::find_slot` to log
information for all streams when debug is enabled.

The motivation for this change is that currently if a non-unified
kv-cache is used, then only one stream will be logged because the
code was currently uses `seq_to_stream[1]`.
This commit is contained in:
Daniel Bevenius
2025-08-11 11:21:19 +02:00
committed by GitHub
parent 50e81bdf5d
commit cd3069dfcb
1 changed files with 56 additions and 52 deletions
Download Patch File Download Diff File Expand all files Collapse all files

108
src/llama-kv-cache-unified.cpp
View File

@@ -738,66 +738,70 @@ bool llama_kv_cache_unified::update(llama_context * lctx, bool do_shift, const d
}
llama_kv_cache_unified::slot_info llama_kv_cache_unified::find_slot(const llama_ubatch & ubatch, bool cont) const {
if (debug > 0) {
const auto & cells = v_cells[seq_to_stream[1]];
for (uint32_t s = 0; s < ubatch.n_seqs_unq; ++s) {
const auto seq_id = ubatch.seq_id_unq[s];
const auto stream_id = seq_to_stream[seq_id];
const auto & cells = v_cells[stream_id];
const uint32_t head_cur = v_heads[stream_id];
const uint32_t head_cur = v_heads[1];
LLAMA_LOG_DEBUG("%s: stream[%d], n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
__func__, stream_id, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
LLAMA_LOG_DEBUG("%s: n = %5d, used = %5d, head = %5d, size = %5d, n_swa = %5d\n",
__func__, cells.used_max_p1(), cells.get_used(), head_cur, get_size(), n_swa);
if ((debug == 2 && n_swa > 0) || debug > 2) {
std::string ss;
for (uint32_t i = 0; i < cells.size(); ++i) {
if (cells.is_empty(i)) {
ss += '.';
} else {
assert(cells.seq_count(i) >= 1);
if (cells.seq_count(i) == 1) {
ss += std::to_string(cells.seq_get(i));
if ((debug == 2 && n_swa > 0) || debug > 2) {
std::string ss;
for (uint32_t i = 0; i < cells.size(); ++i) {
if (cells.is_empty(i)) {
ss += '.';
} else {
ss += 'M';
assert(cells.seq_count(i) >= 1);
if (cells.seq_count(i) == 1) {
ss += std::to_string(cells.seq_get(i));
} else {
ss += 'M';
}
}
if (i%256 == 255) {
ss += " *";
ss += '\n';
}
}
if (i%256 == 255) {
ss += " *";
ss += '\n';
}
}
LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
}
if ((debug == 2 && n_swa > 0) || debug > 2) {
std::string ss;
for (uint32_t i = 0; i < cells.size(); ++i) {
std::string cur;
if (cells.is_empty(i)) {
cur = '.';
} else {
cur = std::to_string(cells.pos_get(i));
}
const int n = cur.size();
for (int j = 0; j < 5 - n; ++j) {
cur += ' ';
}
ss += cur;
if (i%256 == 255) {
ss += " *";
}
if (i%64 == 63) {
ss += '\n';
}
}
LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
}
for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
if (cells.seq_pos_min(s) < 0) {
continue;
LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
}
LLAMA_LOG_DEBUG("%s: min[%d] = %5d, max[%d] = %5d\n", __func__, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
if ((debug == 2 && n_swa > 0) || debug > 2) {
std::string ss;
for (uint32_t i = 0; i < cells.size(); ++i) {
std::string cur;
if (cells.is_empty(i)) {
cur = '.';
} else {
cur = std::to_string(cells.pos_get(i));
}
const int n = cur.size();
for (int j = 0; j < 5 - n; ++j) {
cur += ' ';
}
ss += cur;
if (i%256 == 255) {
ss += " *";
}
if (i%64 == 63) {
ss += '\n';
}
}
LLAMA_LOG_DEBUG("\n%s\n", ss.c_str());
}
for (int s = 0; s < LLAMA_MAX_SEQ; ++s) {
if (cells.seq_pos_min(s) < 0) {
continue;
}
LLAMA_LOG_DEBUG("%s: stream[%d] min[%d] = %5d, max[%d] = %5d\n", __func__, stream_id, s, cells.seq_pos_min(s), s, cells.seq_pos_max(s));
}
}
}