Skip to content
Open
Show file tree
Hide file tree
Changes from all commits
Commits
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
2 changes: 1 addition & 1 deletion src/beam_search_scorer.cpp
Original file line number Diff line number Diff line change
Expand Up @@ -66,7 +66,7 @@ BeamSearchScorer::BeamSearchScorer(const GeneratorParams& parameters)
next_beam_indices_ = parameters.p_device->Allocate<int32_t>(batch_beam_size);

// Space to store intermediate sequence
size_t const per_beam = (max_length_ * (max_length_ + 1)) / 2;
size_t const per_beam = (static_cast<size_t>(max_length_) * (static_cast<size_t>(max_length_) + 1)) / 2;
hypothesis_buffer_ = device.Allocate<int32_t>(batch_beam_size * per_beam);
Comment thread
apsonawane marked this conversation as resolved.

memset(next_beam_scores_.Span().data(), 0, next_beam_scores_.Span().size_bytes());
Expand Down
132 changes: 70 additions & 62 deletions src/config.cpp
Original file line number Diff line number Diff line change
@@ -1,4 +1,4 @@
// Copyright (c) Microsoft Corporation. All rights reserved.
// Copyright (c) Microsoft Corporation. All rights reserved.
// Licensed under the MIT License.
// Modifications Copyright (C) 2026 Advanced Micro Devices, Inc. All rights reserved.
// Portions of this file consist of AI generated content.
Expand Down Expand Up @@ -88,7 +88,7 @@ struct Int_Array_Element : JSON::Element {
explicit Int_Array_Element(std::vector<int>& v) : v_{v} {}

void OnValue(std::string_view name, JSON::Value value) override {
v_.emplace_back(static_cast<int>(JSON::Get<double>(value)));
v_.emplace_back(SafeDoubleToInt(JSON::Get<double>(value), name));
}

private:
Expand Down Expand Up @@ -197,13 +197,13 @@ struct SessionOptions_Element : JSON::Element {
} else if (name == "enable_profiling") {
v_.enable_profiling = JSON::Get<std::string_view>(value);
} else if (name == "intra_op_num_threads") {
v_.intra_op_num_threads = static_cast<int>(JSON::Get<double>(value));
v_.intra_op_num_threads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "inter_op_num_threads") {
v_.inter_op_num_threads = static_cast<int>(JSON::Get<double>(value));
v_.inter_op_num_threads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "log_severity_level") {
v_.log_severity_level = static_cast<int>(JSON::Get<double>(value));
v_.log_severity_level = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "log_verbosity_level") {
v_.log_verbosity_level = static_cast<int>(JSON::Get<double>(value));
v_.log_verbosity_level = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "enable_cpu_mem_arena") {
v_.enable_cpu_mem_arena = JSON::Get<bool>(value);
} else if (name == "enable_mem_pattern") {
Expand Down Expand Up @@ -419,7 +419,7 @@ struct IntArray_Element : JSON::Element {
explicit IntArray_Element(std::vector<int>& v) : v_{v} {}

void OnValue(std::string_view name, JSON::Value value) override {
v_.push_back(static_cast<int>(JSON::Get<double>(value)));
v_.push_back(SafeDoubleToInt(JSON::Get<double>(value), name));
}

private:
Expand Down Expand Up @@ -450,7 +450,7 @@ struct PipelineModel_Element : JSON::Element {
} else if (name == "is_lm_head") {
v_.is_lm_head = JSON::Get<bool>(value);
} else if (name == "reset_session_idx") {
v_.reset_session_idx = static_cast<int>(JSON::Get<double>(value));
v_.reset_session_idx = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -523,9 +523,9 @@ struct SlidingWindow_Element : JSON::Element {

void OnValue(std::string_view name, JSON::Value value) override {
if (name == "window_size") {
v_->window_size = static_cast<int>(JSON::Get<double>(value));
v_->window_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "pad_value") {
v_->pad_value = static_cast<int>(JSON::Get<double>(value));
v_->pad_value = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "alignment") {
v_->alignment = JSON::Get<std::string_view>(value);
} else if (name == "slide_key_value_cache") {
Expand Down Expand Up @@ -560,15 +560,15 @@ struct Encoder_Element : JSON::Element {
if (name == "filename") {
v_.filename = JSON::Get<std::string_view>(value);
} else if (name == "hidden_size") {
v_.hidden_size = static_cast<int>(JSON::Get<double>(value));
v_.hidden_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_attention_heads") {
v_.num_attention_heads = static_cast<int>(JSON::Get<double>(value));
v_.num_attention_heads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_hidden_layers") {
v_.num_hidden_layers = static_cast<int>(JSON::Get<double>(value));
v_.num_hidden_layers = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_key_value_heads") {
v_.num_key_value_heads = static_cast<int>(JSON::Get<double>(value));
v_.num_key_value_heads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "head_size") {
v_.head_size = static_cast<int>(JSON::Get<double>(value));
v_.head_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -609,17 +609,17 @@ struct Decoder_Element : JSON::Element {
if (name == "filename") {
v_.filename = JSON::Get<std::string_view>(value);
} else if (name == "hidden_size") {
v_.hidden_size = static_cast<int>(JSON::Get<double>(value));
v_.hidden_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_attention_heads") {
v_.num_attention_heads = static_cast<int>(JSON::Get<double>(value));
v_.num_attention_heads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_hidden_layers") {
v_.num_hidden_layers = static_cast<int>(JSON::Get<double>(value));
v_.num_hidden_layers = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_key_value_heads") {
v_.num_key_value_heads = static_cast<int>(JSON::Get<double>(value));
v_.num_key_value_heads = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "head_size") {
v_.head_size = static_cast<int>(JSON::Get<double>(value));
v_.head_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "conv_cache_size") {
v_.conv_cache_size = static_cast<int>(JSON::Get<double>(value));
v_.conv_cache_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -795,15 +795,15 @@ struct Vision_Element : JSON::Element {
} else if (name == "adapter_filename") {
v_.adapter_filename = JSON::Get<std::string_view>(value);
} else if (name == "spatial_merge_size") {
v_.spatial_merge_size = static_cast<int>(JSON::Get<double>(value));
v_.spatial_merge_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "tokens_per_second") {
v_.tokens_per_second = static_cast<float>(JSON::Get<double>(value));
} else if (name == "patch_size") {
v_.patch_size = static_cast<int>(JSON::Get<double>(value));
v_.patch_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_visual_tokens") {
v_.num_visual_tokens = static_cast<int>(JSON::Get<double>(value));
v_.num_visual_tokens = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "window_size") {
v_.window_size = static_cast<int>(JSON::Get<double>(value));
v_.window_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -1010,9 +1010,9 @@ struct VAD_Element : JSON::Element {
} else if (name == "threshold") {
v_.threshold = static_cast<float>(JSON::Get<double>(value));
} else if (name == "silence_duration_ms") {
v_.silence_duration_ms = static_cast<int>(JSON::Get<double>(value));
v_.silence_duration_ms = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "prefix_padding_ms") {
v_.prefix_padding_ms = static_cast<int>(JSON::Get<double>(value));
v_.prefix_padding_ms = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -1122,63 +1122,65 @@ struct Model_Element : JSON::Element {
} else if (name == "tokenizer_dir") {
v_.tokenizer_dir = JSON::Get<std::string_view>(value);
} else if (name == "vocab_size") {
v_.vocab_size = static_cast<int>(JSON::Get<double>(value));
v_.vocab_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "context_length") {
v_.context_length = static_cast<int>(JSON::Get<double>(value));
v_.context_length = SafeDoubleToInt(JSON::Get<double>(value), name);
Comment thread
apsonawane marked this conversation as resolved.
if (v_.context_length <= 0)
throw std::out_of_range("context_length must be > 0, got " + std::to_string(v_.context_length));
} else if (name == "pad_token_id") {
Comment thread
apsonawane marked this conversation as resolved.
v_.pad_token_id = static_cast<int>(JSON::Get<double>(value));
v_.pad_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "eos_token_id") {
v_.eos_token_id.assign(1, static_cast<int>(JSON::Get<double>(value)));
v_.eos_token_id.assign(1, SafeDoubleToInt(JSON::Get<double>(value), name));
} else if (name == "bos_token_id") {
v_.bos_token_id = static_cast<int>(JSON::Get<double>(value));
v_.bos_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "decoder_start_token_id") {
v_.decoder_start_token_id = static_cast<int>(JSON::Get<double>(value));
v_.decoder_start_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "sep_token_id") {
v_.sep_token_id = static_cast<int>(JSON::Get<double>(value));
v_.sep_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "image_token_id") {
v_.image_token_id = static_cast<int>(JSON::Get<double>(value));
v_.image_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "audio_token_id") {
v_.audio_token_id = static_cast<int>(JSON::Get<double>(value));
v_.audio_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "boa_token_id") {
v_.boa_token_id = static_cast<int>(JSON::Get<double>(value));
v_.boa_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "video_token_id") {
v_.video_token_id = static_cast<int>(JSON::Get<double>(value));
v_.video_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "vision_start_token_id") {
v_.vision_start_token_id = static_cast<int>(JSON::Get<double>(value));
v_.vision_start_token_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_mels") {
v_.num_mels = static_cast<int>(JSON::Get<double>(value));
v_.num_mels = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "fft_size") {
v_.fft_size = static_cast<int>(JSON::Get<double>(value));
v_.fft_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "hop_length") {
v_.hop_length = static_cast<int>(JSON::Get<double>(value));
v_.hop_length = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "win_length") {
v_.win_length = static_cast<int>(JSON::Get<double>(value));
v_.win_length = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "preemph") {
v_.preemph = static_cast<float>(JSON::Get<double>(value));
} else if (name == "log_eps") {
v_.log_eps = static_cast<float>(JSON::Get<double>(value));
} else if (name == "norm_eps") {
v_.norm_eps = static_cast<float>(JSON::Get<double>(value));
} else if (name == "subsampling_factor") {
v_.subsampling_factor = static_cast<int>(JSON::Get<double>(value));
v_.subsampling_factor = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "left_context") {
v_.left_context = static_cast<int>(JSON::Get<double>(value));
v_.left_context = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "conv_context") {
v_.conv_context = static_cast<int>(JSON::Get<double>(value));
v_.conv_context = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "pre_encode_cache_size") {
v_.pre_encode_cache_size = static_cast<int>(JSON::Get<double>(value));
v_.pre_encode_cache_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "sample_rate") {
v_.sample_rate = static_cast<int>(JSON::Get<double>(value));
v_.sample_rate = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "chunk_samples") {
v_.chunk_samples = static_cast<int>(JSON::Get<double>(value));
v_.chunk_samples = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "blank_id") {
v_.blank_id = static_cast<int>(JSON::Get<double>(value));
v_.blank_id = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "max_symbols_per_step") {
v_.max_symbols_per_step = static_cast<int>(JSON::Get<double>(value));
v_.max_symbols_per_step = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "left_context_samples") {
v_.left_context_samples = static_cast<int>(JSON::Get<double>(value));
v_.left_context_samples = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "right_context_samples") {
v_.right_context_samples = static_cast<int>(JSON::Get<double>(value));
v_.right_context_samples = SafeDoubleToInt(JSON::Get<double>(value), name);
} else {
throw JSON::unknown_value_error{};
}
Expand Down Expand Up @@ -1249,8 +1251,14 @@ int SafeDoubleToInt(double x, std::string_view name) {
throw std::overflow_error(ss.str());
}

// 3. Perform the cast. This truncates any fractional part (e.g., 3.9 becomes 3).
// If rounding is desired, use `return static_cast<int>(std::round(x));`
// 3. Reject fractional values — these fields must be integral.
if (x != std::trunc(x)) {
std::stringstream ss;
ss << "Field '" << name << "' value " << x << " is not an integer";
throw std::invalid_argument(ss.str());
}

// 4. Perform the cast.
return static_cast<int>(x);
}

Expand All @@ -1259,17 +1267,17 @@ struct Search_Element : JSON::Element {

void OnValue(std::string_view name, JSON::Value value) override {
if (name == "min_length") {
v_.min_length = static_cast<int>(JSON::Get<double>(value));
v_.min_length = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "max_length") {
v_.max_length = static_cast<int>(JSON::Get<double>(value));
v_.max_length = SafeDoubleToInt(JSON::Get<double>(value), name);
Comment thread
apsonawane marked this conversation as resolved.
} else if (name == "batch_size") {
v_.batch_size = static_cast<int>(JSON::Get<double>(value));
v_.batch_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_beams") {
Comment thread
apsonawane marked this conversation as resolved.
v_.num_beams = static_cast<int>(JSON::Get<double>(value));
v_.num_beams = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "num_return_sequences") {
v_.num_return_sequences = static_cast<int>(JSON::Get<double>(value));
v_.num_return_sequences = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "top_k") {
v_.top_k = static_cast<int>(JSON::Get<double>(value));
v_.top_k = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "top_p") {
v_.top_p = static_cast<float>(JSON::Get<double>(value));
} else if (name == "temperature") {
Expand All @@ -1279,7 +1287,7 @@ struct Search_Element : JSON::Element {
} else if (name == "length_penalty") {
v_.length_penalty = static_cast<float>(JSON::Get<double>(value));
} else if (name == "no_repeat_ngram_size") {
v_.no_repeat_ngram_size = static_cast<int>(JSON::Get<double>(value));
v_.no_repeat_ngram_size = SafeDoubleToInt(JSON::Get<double>(value), name);
} else if (name == "diversity_penalty") {
v_.diversity_penalty = static_cast<float>(JSON::Get<double>(value));
} else if (name == "length_penalty") {
Expand Down
2 changes: 2 additions & 0 deletions src/config.h
Original file line number Diff line number Diff line change
Expand Up @@ -465,6 +465,8 @@ void SetSearchBool(Config::Search& search, std::string_view name, bool value);
void ClearProviders(Config& config);
void SetProviderOption(Config& config, std::string_view provider_name, std::string_view option_name, std::string_view option_value);
void OverlayConfig(Config& config, std::string_view json);
int SafeDoubleToInt(double x, std::string_view name);

bool IsGraphCaptureEnabled(const Config::SessionOptions& session_options);
bool IsMultiProfileEnabled(const Config::SessionOptions& session_options);

Expand Down
Loading