Skip to content

feat: add sequence parallel and cache support for Flux DiT. #2236

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
z-jun03 wants to merge 1 commit into
base: main
Choose a base branch
from
+373 −55
Open

feat: add sequence parallel and cache support for Flux DiT. #2236

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
13 changes: 10 additions & 3 deletions xllm/models/dit/pipelines/pipeline_flux.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,8 @@ class FluxPipelineImpl : public FluxPipelineBaseImpl {
"pos_embed",
FluxPosEmbed(ROPE_SCALE_BASE,
context.get_model_args("transformer").axes_dims_rope()));
transformer_ = FluxDiTModel(context.get_model_context("transformer"));
transformer_ = FluxDiTModel(context.get_model_context("transformer"),
context.get_parallel_args());
t5_ = T5EncoderModel(context.get_model_context("text_encoder_2"));
clip_text_model_ = CLIPTextModel(context.get_model_context("text_encoder"));
scheduler_ =
Expand Down Expand Up @@ -275,6 +276,10 @@ class FluxPipelineImpl : public FluxPipelineBaseImpl {
torch::Tensor image_rotary_emb =
torch::stack({rot_emb1, rot_emb2}, 0).to(options_.dtype());

DiTCache::get_instance().set_context(
{/*infer_steps=*/static_cast<int64_t>(timesteps.numel()),
/*num_blocks=*/transformer_->num_blocks()});

for (int64_t i = 0; i < timesteps.numel(); ++i) {
torch::Tensor t = timesteps[i].unsqueeze(0);
timestep.fill_(t.item<float>())
Expand All @@ -287,7 +292,8 @@ class FluxPipelineImpl : public FluxPipelineBaseImpl {
timestep,
image_rotary_emb,
guidance,
step_id);
step_id,
/*use_cfg=*/false);
if (do_true_cfg) {
torch::Tensor negative_noise_pred =
transformer_->forward(prepared_latents,
Expand All @@ -296,7 +302,8 @@ class FluxPipelineImpl : public FluxPipelineBaseImpl {
timestep,
image_rotary_emb,
guidance,
step_id);
step_id,
/*use_cfg=*/true);
noise_pred =
noise_pred + (noise_pred - negative_noise_pred) * true_cfg_scale;
negative_noise_pred.reset();
Expand Down
6 changes: 5 additions & 1 deletion xllm/models/dit/pipelines/pipeline_flux_control.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -46,7 +46,8 @@ class FluxControlPipelineImpl : public FluxPipelineBaseImpl {
"pos_embed",
FluxPosEmbed(ROPE_SCALE_BASE,
context.get_model_args("transformer").axes_dims_rope()));
transformer_ = FluxDiTModel(context.get_model_context("transformer"));
transformer_ = FluxDiTModel(context.get_model_context("transformer"),
context.get_parallel_args());
t5_ = T5EncoderModel(context.get_model_context("text_encoder_2"));
clip_text_model_ = CLIPTextModel(context.get_model_context("text_encoder"));
scheduler_ =
Expand Down Expand Up @@ -298,6 +299,9 @@ class FluxControlPipelineImpl : public FluxPipelineBaseImpl {
width / (vae_scale_factor_ * 2));
torch::Tensor image_rotary_emb =
torch::stack({rot_emb1, rot_emb2}, 0).to(options_.dtype());
DiTCache::get_instance().set_context(
{/*infer_steps=*/static_cast<int64_t>(timesteps.numel()),
/*num_blocks=*/transformer_->num_blocks()});
for (int64_t i = 0; i < timesteps.numel(); ++i) {
torch::Tensor t = timesteps[i].unsqueeze(0);
timestep.fill_(t.item<float>())
Expand Down
7 changes: 6 additions & 1 deletion xllm/models/dit/pipelines/pipeline_flux_fill.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -54,7 +54,8 @@ class FluxFillPipelineImpl : public FluxPipelineBaseImpl {
"pos_embed",
FluxPosEmbed(ROPE_SCALE_BASE,
context.get_model_args("transformer").axes_dims_rope()));
transformer_ = FluxDiTModel(context.get_model_context("transformer"));
transformer_ = FluxDiTModel(context.get_model_context("transformer"),
context.get_parallel_args());
t5_ = T5EncoderModel(context.get_model_context("text_encoder_2"));
clip_text_model_ = CLIPTextModel(context.get_model_context("text_encoder"));
scheduler_ =
Expand Down Expand Up @@ -378,6 +379,10 @@ class FluxFillPipelineImpl : public FluxPipelineBaseImpl {
torch::Tensor image_rotary_emb =
torch::stack({rot_emb1, rot_emb2}, 0).to(options_.dtype());

DiTCache::get_instance().set_context(
{/*infer_steps=*/static_cast<int64_t>(timesteps.size(0)),
/*num_blocks=*/transformer_->num_blocks()});

for (int64_t i = 0; i < timesteps.size(0); ++i) {
torch::Tensor t = timesteps[i];
torch::Tensor timestep =
Expand Down
Loading
Loading