Skip to content

perplexity: Add option to ignore context window overflow errors and continue score calculation #12512

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 9 commits into from
Closed

perplexity: Add option to ignore context window overflow errors and continue score calculation #12512

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
41a63f4
Ensure tasks are randomized across test runs
EAddario Mar 15, 2025
7fbbadc
Error log offending task
EAddario Mar 15, 2025
fa7d1c6
Add --ignore-context-overflow option for llama-perplexity
EAddario Mar 15, 2025
1035c6e
Add option to ignore context overflow during score computation
EAddario Mar 15, 2025
8a827f8
Merge branch 'master' into perplexity
EAddario Mar 16, 2025
f348628
Merge branch 'master' into perplexity
EAddario Mar 19, 2025
e354bc3
Merge branch 'master' into perplexity
EAddario Mar 22, 2025
d41f015
Merge branch 'master' into perplexity
EAddario Mar 29, 2025
c5cd061
Merge branch 'master' into perplexity
EAddario Apr 7, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 7 additions & 0 deletions common/arg.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2030,6 +2030,13 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
params.multiple_choice_tasks = value;
}
).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
add_opt(common_arg(
{"--ignore-context-overflow"},
string_format("ignores context window overflow when computing scores (default: %s)", params.ctx_overflow ? "enabled" : "disabled"),
[](common_params & params) {
params.ctx_overflow = true;
}
).set_examples({LLAMA_EXAMPLE_PERPLEXITY}));
add_opt(common_arg(
{"--kl-divergence"},
"computes KL-divergence to logits provided via --kl-divergence-base",
Expand Down
2 changes: 2 additions & 0 deletions common/common.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -303,6 +303,8 @@ struct common_params {
bool multiple_choice = false; // compute TruthfulQA score over random tasks from datafile supplied in prompt
size_t multiple_choice_tasks = 0; // number of tasks to use when computing the TruthfulQA score. If 0, all tasks will be computed

bool ctx_overflow = false; // ignore context window overflow during HellaSwag, WinoGrande or Multiple Choice evaluation

bool kl_divergence = false; // compute KL divergence

bool usage = false; // print usage
Expand Down
22 changes: 18 additions & 4 deletions examples/perplexity/perplexity.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -786,8 +786,7 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {
hs_task_count = params.hellaswag_tasks;
}

// The random seed should not impact the final result if the computation is done over enough tasks, so kept hardcoded for now
std::mt19937 rng(1);
std::mt19937 rng(std::random_device{}());

// Dataholder for hellaswag tasks
struct hs_data_t {
Expand Down Expand Up @@ -921,6 +920,11 @@ static void hellaswag_score(llama_context * ctx, const common_params & params) {

if (i0 == i1) {
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
if (params.ctx_overflow) {
LOG_ERR("%s : ignoring offending task: %s\n", __func__, hs_data[i0].context.c_str());
continue;
}

return;
}

Expand Down Expand Up @@ -1111,7 +1115,7 @@ static void winogrande_score(llama_context * ctx, const common_params & params)

if (params.winogrande_tasks > 0 && params.winogrande_tasks < data.size()) {
LOG_INF("%s : selecting %zu random tasks\n", __func__, params.winogrande_tasks);
std::mt19937 rng(1);
std::mt19937 rng(std::random_device{}());
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I suspect that using a fixed seed was intentional so that the results are repeatable.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I thought so too, but the rationale for the proposed change is that by "freezing" the test cases it lessens the validity of the test. If repeatable results are indeed the objective, why go through the extra trouble of randomizing in the first place, instead of just processing tests cases sequentially?

I realise this is more of a personal preference and I'm OK to revert the change back if needed.

std::vector<int> aux(data.size());
for (int i = 0; i < int(data.size()); ++i) {
aux[i] = i;
Expand Down Expand Up @@ -1214,6 +1218,11 @@ static void winogrande_score(llama_context * ctx, const common_params & params)

if (i0 == i1) {
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
if (params.ctx_overflow) {
LOG_ERR("%s : offending task: %s\n", __func__, data[i0].first.c_str());
continue;
}

return;
}

Expand Down Expand Up @@ -1437,7 +1446,7 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par
}
else {
LOG_INF("%s: selecting %zu random tasks from %u tasks available\n", __func__, params.multiple_choice_tasks, n_task);
std::mt19937 rng(1);
std::mt19937 rng(std::random_device{}());
std::vector<int> aux(n_task);
for (uint32_t i = 0; i < n_task; ++i) aux[i] = i;
float scale = 1.f/(1.f + (float)std::mt19937::max());
Expand Down Expand Up @@ -1586,6 +1595,11 @@ static void multiple_choice_score(llama_context * ctx, const common_params & par

if (i0 == i1) {
LOG_ERR("%s : task %zu does not fit in the context window\n", __func__, i0);
if (params.ctx_overflow) {
LOG_ERR("%s : offending task: %s\n", __func__, tasks[i0].question.c_str());
continue;
}

return;
}

Expand Down