Skip to content

refactor: split forward inputs from model input params [3 / 3]. #1469

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

Open
RobbieLeung wants to merge 1 commit into
base: main
Choose a base branch
from
Open

refactor: split forward inputs from model input params [3 / 3]. #1469

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
125 changes: 64 additions & 61 deletions tests/core/framework/batch/batch_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -324,47 +324,48 @@ TEST(BatchTest, Basic) {
EXPECT_TRUE(equal(forward_input.positions, expected_pos));

// check the input parameters
const ModelInputParams& input_params = forward_input.input_params;
EXPECT_TRUE(input_params.meta.batch_forward_type.is_mixed());
EXPECT_EQ(input_params.meta.num_sequences, 4);
EXPECT_EQ(input_params.meta.q_max_seq_len, 9);
EXPECT_EQ(input_params.meta.kv_max_seq_len, 16);
EXPECT_EQ(input_params.embedding.embedding_ids, std::vector<int32_t>({-1, -1, -1}));
EXPECT_EQ(input_params.embedding.linear_state_ids,
EXPECT_TRUE(forward_input.meta.batch_forward_type.is_mixed());
EXPECT_EQ(forward_input.meta.num_sequences, 4);
EXPECT_EQ(forward_input.meta.q_max_seq_len, 9);
EXPECT_EQ(forward_input.meta.kv_max_seq_len, 16);
EXPECT_EQ(forward_input.embedding.embedding_ids,
std::vector<int32_t>({-1, -1, -1}));
EXPECT_EQ(forward_input.embedding.linear_state_ids,
std::vector<int32_t>({-1, -1, -1, -1}));

#if defined(USE_NPU)
const std::vector<int32_t> q_seq_lens = {9, 1, 1, 4};
#else
const std::vector<int32_t> q_seq_lens = {0, 9, 10, 11, 15};
#endif
EXPECT_TRUE(equal(input_params.attention.device.q_seq_lens, q_seq_lens));
EXPECT_TRUE(equal(forward_input.attention.device.q_seq_lens, q_seq_lens));

// seq4's kv_seq_len = q_len + num_cached_tokens (q_len<=max_allowed_tokens)
#if defined(USE_NPU)
const std::vector<int32_t> kv_seq_lens = {9, 8, 16, 8};
#else
const std::vector<int32_t> kv_seq_lens = {0, 9, 17, 33, 41};
#endif
EXPECT_TRUE(equal(input_params.attention.device.kv_seq_lens, kv_seq_lens));
EXPECT_TRUE(equal(forward_input.attention.device.kv_seq_lens, kv_seq_lens));

const std::vector<int32_t> new_cache_slots = {
/*seq1*/ 4, 5, 6, 7, 8, 9, 10, 11, 12,
/*seq2*/ 23,
/*seq3*/ 47,
/*seq4*/ 56,57,58,59
};
EXPECT_TRUE(equal(input_params.attention.device.new_cache_slots, new_cache_slots));
EXPECT_TRUE(equal(forward_input.attention.device.new_cache_slots,
new_cache_slots));

const std::vector<int32_t> block_tables = {
/*seq1*/ 1, 2, 3, 0, 0,
/*seq2*/ 4, 5, 6, 7, 0,
/*seq3*/ 8, 9, 10, 11, 12,
/*seq4*/ 13, 14, 15, 0, 0};
EXPECT_TRUE(equal(input_params.attention.device.block_tables, block_tables));
EXPECT_TRUE(equal(forward_input.attention.device.block_tables, block_tables));

// const std::vector<int32_t> last_token_idxes = {8, 9, 10};
// EXPECT_TRUE(equal(input_params.last_token_idxes, last_token_idxes));
// EXPECT_TRUE(equal(forward_input.last_token_idxes, last_token_idxes));

const auto& sampling_params = forward_input.sampling_params;
const std::vector<int64_t> unique_ids = {
Expand Down Expand Up @@ -578,7 +579,7 @@ TEST(BatchTest, ForwardInputPreservesTransferInfoAndBatchId) {
(std::vector<uint64_t>{1, 2}));
EXPECT_EQ(input.transfer_kv_infos[0].remote_blocks_ids,
(std::vector<uint64_t>{100, 101}));
EXPECT_EQ(input.input_params.meta.batch_id, batch_id);
EXPECT_EQ(input.meta.batch_id, batch_id);
}

TEST(BatchTest, ForwardInputPackedRoundTripPreservesTransportFields) {
Expand Down Expand Up @@ -650,7 +651,7 @@ TEST(BatchTest, ForwardInputPackedRoundTripPreservesTransportFields) {
ForwardInput round_trip;
reader_manager.input_read(round_trip, torch::Device(torch::kCPU));

EXPECT_EQ(round_trip.input_params.meta.batch_id, batch_id);
EXPECT_EQ(round_trip.meta.batch_id, batch_id);
EXPECT_TRUE(equal(round_trip.token_ids, std::vector<int32_t>({1, 2, 3, 4})));
ASSERT_EQ(round_trip.transfer_kv_infos.size(), 1u);
EXPECT_EQ(round_trip.transfer_kv_infos[0].local_blocks_ids,
Expand Down Expand Up @@ -709,18 +710,18 @@ TEST(BatchTest, ForwardInputBlockCopyKernelFieldsMatchExpectedLayout) {
forward_builder.build_forward_input(/*num_decoding_tokens=*/1,
/*min_decoding_batch_size=*/0);

EXPECT_TRUE(equal(forward_input.input_params.block_copy.src_block_indices,
EXPECT_TRUE(equal(forward_input.block_copy.src_block_indices,
std::vector<int32_t>({7, 8})));
EXPECT_TRUE(equal(forward_input.input_params.block_copy.dst_block_indices,
EXPECT_TRUE(equal(forward_input.block_copy.dst_block_indices,
std::vector<int32_t>({10, 11, 12})));
EXPECT_TRUE(equal(forward_input.input_params.block_copy.cum_sum,
std::vector<int32_t>({2, 3})));
EXPECT_TRUE(
equal(forward_input.block_copy.cum_sum, std::vector<int32_t>({2, 3})));

#if defined(USE_CUDA)
EXPECT_EQ(forward_input.input_params.block_copy.swap_blocks.size(),
EXPECT_EQ(forward_input.block_copy.swap_blocks.size(),
forward_swap_blocks.size());
#else
EXPECT_TRUE(forward_input.input_params.block_copy.swap_blocks.empty());
EXPECT_TRUE(forward_input.block_copy.swap_blocks.empty());
#endif

FLAGS_enable_block_copy_kernel = old_enable_block_copy_kernel;
Expand Down Expand Up @@ -780,13 +781,13 @@ TEST(BatchTest, ForwardInputCpPartitionMatchesExpectedLayout) {
equal(cp_forward_input.positions, std::vector<int32_t>({0, 1, 6, 7})));
EXPECT_TRUE(equal(cp_forward_input.sampling_params.selected_token_idxes,
std::vector<int32_t>({3})));
EXPECT_EQ(cp_forward_input.input_params.meta.q_max_seq_len, 4);
EXPECT_EQ(cp_forward_input.input_params.meta.kv_max_seq_len, 4);
EXPECT_EQ(cp_forward_input.meta.q_max_seq_len, 4);
EXPECT_EQ(cp_forward_input.meta.kv_max_seq_len, 4);

const std::vector<int32_t>& q_seq_lens =
cp_forward_input.input_params.attention.host.q_seq_lens;
cp_forward_input.attention.host.q_seq_lens;
const std::vector<int32_t>& kv_seq_lens =
cp_forward_input.input_params.attention.host.kv_seq_lens;
cp_forward_input.attention.host.kv_seq_lens;
EXPECT_TRUE((q_seq_lens == std::vector<int32_t>({4}) ||
q_seq_lens == std::vector<int32_t>({0, 4})));
EXPECT_TRUE((kv_seq_lens == std::vector<int32_t>({4}) ||
Expand Down Expand Up @@ -910,13 +911,11 @@ TEST(BatchTest, SampleRequestKeepsThreadedForwardBuilderOffsetsStable) {
expected_selected_token_idxes));
EXPECT_TRUE(
equal(forward_input.sampling_params.sample_idxes, expected_sample_idxes));
ASSERT_EQ(forward_input.input_params.embedding.embedding_ids.size(),
sequences.size());
ASSERT_EQ(forward_input.input_params.embedding.linear_state_ids.size(),
sequences.size());
EXPECT_EQ(forward_input.input_params.embedding.embedding_ids,
ASSERT_EQ(forward_input.embedding.embedding_ids.size(), sequences.size());
ASSERT_EQ(forward_input.embedding.linear_state_ids.size(), sequences.size());
EXPECT_EQ(forward_input.embedding.embedding_ids,
std::vector<int32_t>({-1, -1}));
EXPECT_EQ(forward_input.input_params.embedding.linear_state_ids,
EXPECT_EQ(forward_input.embedding.linear_state_ids,
std::vector<int32_t>({-1, -1}));
}

Expand Down Expand Up @@ -967,11 +966,10 @@ TEST(BatchTest, DecodeMinBatchSizeDoesNotPadTransportState) {
builder.build_forward_input(/*num_decoding_tokens=*/1,
/*min_decoding_batch_size=*/3);

EXPECT_EQ(forward_input.input_params.meta.num_sequences, 1);
EXPECT_EQ(forward_input.input_params.embedding.linear_state_ids,
std::vector<int32_t>({-1}));
EXPECT_EQ(forward_input.input_params.embedding.embedding_ids,
EXPECT_EQ(forward_input.meta.num_sequences, 1);
EXPECT_EQ(forward_input.embedding.linear_state_ids,
std::vector<int32_t>({-1}));
EXPECT_EQ(forward_input.embedding.embedding_ids, std::vector<int32_t>({-1}));
}

TEST(BatchTest, DecodeSingleBlockIdsStaySplitInTransportButShareSlotValue) {
Expand Down Expand Up @@ -1026,15 +1024,13 @@ TEST(BatchTest, DecodeSingleBlockIdsStaySplitInTransportButShareSlotValue) {
builder.build_forward_input(/*num_decoding_tokens=*/1,
/*min_decoding_batch_size=*/0);

ASSERT_EQ(forward_input.input_params.embedding.embedding_ids.size(), 1u);
ASSERT_EQ(forward_input.input_params.embedding.linear_state_ids.size(), 1u);
EXPECT_EQ(forward_input.input_params.embedding.embedding_ids[0],
expected_slot_id);
EXPECT_EQ(forward_input.input_params.embedding.linear_state_ids[0],
expected_slot_id);
ASSERT_EQ(forward_input.embedding.embedding_ids.size(), 1u);
ASSERT_EQ(forward_input.embedding.linear_state_ids.size(), 1u);
EXPECT_EQ(forward_input.embedding.embedding_ids[0], expected_slot_id);
EXPECT_EQ(forward_input.embedding.linear_state_ids[0], expected_slot_id);
}

TEST(BatchTest, SharedMemoryRoundTripPreservesLinearStateIds) {
TEST(BatchTest, SharedMemoryRoundTripPreservesAndDefaultsLinearStateIds) {
ForwardInput forward_input;
auto int_options = torch::TensorOptions()
.dtype(torch::kInt)
Expand All @@ -1046,30 +1042,30 @@ TEST(BatchTest, SharedMemoryRoundTripPreservesLinearStateIds) {
forward_input.positions =
torch::tensor(std::vector<int32_t>({0, 0}), int_options);
forward_input.positions_host = forward_input.positions;
forward_input.input_params.meta.batch_forward_type = BatchForwardType::DECODE;
forward_input.input_params.meta.num_sequences = 2;
forward_input.input_params.meta.kv_max_seq_len = 1;
forward_input.input_params.meta.q_max_seq_len = 1;
forward_input.input_params.attention.host.kv_seq_lens = {1, 1};
forward_input.input_params.attention.host.q_seq_lens = {1, 1};
forward_input.input_params.attention.host.q_cu_seq_lens = {1, 2};
forward_input.input_params.attention.host.kv_cache_tokens_nums = {0, 0};
forward_input.input_params.attention.host.new_cache_slots = {0, 0};
forward_input.input_params.attention.device.kv_seq_lens =
forward_input.meta.batch_forward_type = BatchForwardType::DECODE;
forward_input.meta.num_sequences = 2;
forward_input.meta.kv_max_seq_len = 1;
forward_input.meta.q_max_seq_len = 1;
forward_input.attention.host.kv_seq_lens = {1, 1};
forward_input.attention.host.q_seq_lens = {1, 1};
forward_input.attention.host.q_cu_seq_lens = {1, 2};
forward_input.attention.host.kv_cache_tokens_nums = {0, 0};
forward_input.attention.host.new_cache_slots = {0, 0};
forward_input.attention.device.kv_seq_lens =
torch::tensor(std::vector<int32_t>({1, 1}), int_options);
forward_input.input_params.attention.device.q_seq_lens =
forward_input.attention.device.q_seq_lens =
torch::tensor(std::vector<int32_t>({1, 1}), int_options);
forward_input.input_params.attention.device.q_cu_seq_lens =
forward_input.attention.device.q_cu_seq_lens =
torch::tensor(std::vector<int32_t>({1, 2}), int_options);
forward_input.input_params.attention.device.kv_cache_tokens_nums =
forward_input.attention.device.kv_cache_tokens_nums =
torch::tensor(std::vector<int32_t>({0, 0}), int_options);
forward_input.input_params.attention.device.new_cache_slots =
forward_input.attention.device.new_cache_slots =
torch::tensor(std::vector<int32_t>({0, 0}), int_options);
forward_input.input_params.attention.device.block_tables = create_2d_tensor(
forward_input.attention.device.block_tables = create_2d_tensor(
std::vector<std::vector<int32_t>>{{0}, {0}}, torch::kInt);
forward_input.input_params.attention.host.block_tables =
forward_input.input_params.attention.device.block_tables;
forward_input.input_params.embedding.linear_state_ids = {4, 6};
forward_input.attention.host.block_tables =
forward_input.attention.device.block_tables;
forward_input.embedding.linear_state_ids = {4, 6};

bool is_creator = false;
auto shm_name =
Expand All @@ -1086,8 +1082,15 @@ TEST(BatchTest, SharedMemoryRoundTripPreservesLinearStateIds) {

ForwardInput from_shm;
reader_manager.input_read(from_shm, torch::Device(torch::kCPU));
EXPECT_EQ(from_shm.input_params.embedding.linear_state_ids,
std::vector<int32_t>({4, 6}));
EXPECT_EQ(from_shm.embedding.linear_state_ids, std::vector<int32_t>({4, 6}));

forward_input.embedding.linear_state_ids.clear();
ASSERT_TRUE(writer_manager.input_write(forward_input));

ForwardInput legacy_from_shm;
reader_manager.input_read(legacy_from_shm, torch::Device(torch::kCPU));
EXPECT_EQ(legacy_from_shm.embedding.linear_state_ids,
std::vector<int32_t>({-1, -1}));
}

TEST(BatchTest, SampleRequestProcessesAllMatchedRawOutputs) {
Expand Down
10 changes: 3 additions & 7 deletions tests/core/framework/hf_model_loader_test.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -41,14 +41,10 @@ class DummyRecCausalLM final : public RecCausalLM {
explicit DummyRecCausalLM(const torch::TensorOptions& options)
: options_(options) {}

ModelOutput forward(const torch::Tensor& tokens,
const torch::Tensor& positions,
std::vector<KVCache>& kv_caches,
const ModelInputParams& parameters) override {
UNUSED_PARAMETER(tokens);
UNUSED_PARAMETER(positions);
ModelOutput forward(const ForwardInput& input,
std::vector<KVCache>& kv_caches) override {
UNUSED_PARAMETER(input);
UNUSED_PARAMETER(kv_caches);
UNUSED_PARAMETER(parameters);
return ModelOutput();
}

Expand Down
Loading
Loading