Skip to content
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

Push rotary embeddings as row-major for e2e speedup #13636

Closed
wants to merge 1 commit into from
Closed

Push rotary embeddings as row-major for e2e speedup #13636

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
1 change: 0 additions & 1 deletion models/demos/t3000/llama2_70b/tests/test_llama_attention.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -194,7 +194,6 @@ def tt_llama_attention_prepare_inputs(llama_attention_model, x, start_pos, mode)
mesh_mapper=ReplicateTensorToMesh(llama_attention_model.mesh_device),
device=llama_attention_model.mesh_device,
)
rot_mats = ttnn.to_device(rot_mats, llama_attention_model.mesh_device)

rot_mats = ttnn.interleaved_to_sharded(rot_mats, llama_attention_model.model_config["ROT_MAT_MM_IN1_MEMCFG"])

Expand Down
1 change: 0 additions & 1 deletion models/demos/t3000/llama2_70b/tests/test_llama_decoder.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -188,7 +188,6 @@ def tt_llama_decoder_prepare_inputs(llama_decoder_model, x, start_pos, mode):
mesh_mapper=ReplicateTensorToMesh(llama_decoder_model.mesh_device),
device=llama_decoder_model.mesh_device,
)
rot_mats = ttnn.to_device(rot_mats, llama_decoder_model.mesh_device)

rot_mats = ttnn.interleaved_to_sharded(rot_mats, llama_decoder_model.model_config["ROT_MAT_MM_IN1_MEMCFG"])

Expand Down
3 changes: 3 additions & 0 deletions models/demos/t3000/llama2_70b/tests/test_llama_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,6 +165,9 @@ def run_test_LlamaModel_inference(
if mode == "decode":
tt_inp_emb = ttnn.to_device(tt_inp_emb, t3k_mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
tt_inp_emb = tt_model.tt_embd(tt_inp_emb)
rot_mat_rm = ttnn.to_device(rot_mat, t3k_mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
rot_mat = ttnn.to_layout(rot_mat_rm, ttnn.TILE_LAYOUT)
rot_mat = ttnn.interleaved_to_sharded(rot_mat, model_config["ROT_MAT_MM_IN1_MEMCFG"])
tt_inp_emb = ttnn.interleaved_to_sharded(tt_inp_emb, model_config["WORD_EMBEDDING_OUTPUT_MEMCFG"])
rot_mat = ttnn.to_device(rot_mat, t3k_mesh_device, memory_config=model_config["ROT_MAT_MM_IN1_MEMCFG"])
cache_idxs = ttnn.to_device(cache_idxs, t3k_mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
Expand Down
4 changes: 3 additions & 1 deletion models/demos/t3000/llama2_70b/tests/test_llama_perf_decode.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -132,7 +132,9 @@ def run_test_LlamaModel_end_to_end(
tt_inp_emb = tt_model.tt_embd(tt_inp_emb)
tt_inp_emb = ttnn.interleaved_to_sharded(tt_inp_emb, tt_model.model_config["WORD_EMBEDDING_OUTPUT_MEMCFG"])

rot_mat = ttnn.to_device(rot_mat, mesh_device, memory_config=tt_model.model_config["ROT_MAT_MM_IN1_MEMCFG"])
rot_mat_rm = ttnn.to_device(rot_mat, mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
rot_mat = ttnn.to_layout(rot_mat_rm, ttnn.TILE_LAYOUT)
rot_mat = ttnn.interleaved_to_sharded(rot_mat, model_config["ROT_MAT_MM_IN1_MEMCFG"])
cache_idxs = ttnn.to_device(cache_idxs, mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)

##### Compile Model #####
Expand Down
8 changes: 6 additions & 2 deletions models/demos/t3000/llama2_70b/tt/llama_generation.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -124,7 +124,9 @@ def capture_trace(self, tokens: torch.Tensor, start_pos: int):
tt_inp = ttnn.to_device(tt_inp, self.mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
tt_inp_emb = self.tt_model.tt_embd(tt_inp)
tt_inp_emb = ttnn.interleaved_to_sharded(tt_inp_emb, self.model_config["WORD_EMBEDDING_OUTPUT_MEMCFG"])
rot_mat = ttnn.to_device(rot_mat, self.mesh_device, memory_config=self.model_config["ROT_MAT_MM_IN1_MEMCFG"])
rot_mat_rm = ttnn.to_device(rot_mat, self.mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
rot_mat = ttnn.to_layout(rot_mat_rm, ttnn.TILE_LAYOUT)
rot_mat = ttnn.interleaved_to_sharded(rot_mat, self.model_config["ROT_MAT_MM_IN1_MEMCFG"])
cache_idxs_tt = ttnn.to_device(cache_idxs_tt, self.mesh_device, memory_config=ttnn.DRAM_MEMORY_CONFIG)
tt_logits = self.tt_model(tt_inp_emb, rot_mat, start_pos, cache_idxs=cache_idxs_tt, mode="decode")

Expand All @@ -134,12 +136,14 @@ def capture_trace(self, tokens: torch.Tensor, start_pos: int):
# Run TT model
tt_inp_emb = self.tt_model.tt_embd(tt_inp)
tt_inp_emb = ttnn.interleaved_to_sharded(tt_inp_emb, self.model_config["WORD_EMBEDDING_OUTPUT_MEMCFG"])
rot_mat = ttnn.to_layout(rot_mat_rm, ttnn.TILE_LAYOUT)
rot_mat = ttnn.interleaved_to_sharded(rot_mat, self.model_config["ROT_MAT_MM_IN1_MEMCFG"])
tt_logits = self.tt_model(tt_inp_emb, rot_mat, start_pos, cache_idxs=cache_idxs_tt, mode="decode")

ttnn.end_trace_capture(self.mesh_device, trace_id, cq_id=0)
logger.info("Done Capturing Decode Trace")

return trace_id, tt_inp, rot_mat, cache_idxs_tt, tt_logits
return trace_id, tt_inp, rot_mat_rm, cache_idxs_tt, tt_logits

def delete_trace(self, trace_id):
ttnn.release_trace(self.mesh_device, trace_id)
Expand Down
2 changes: 1 addition & 1 deletion models/demos/t3000/llama2_70b/tt/llama_model_optimized.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -258,7 +258,7 @@ def prepare_inputs(self, inp_ids, start_pos, valid_seq_len=None, mode="decode"):
rot_mats = ttnn.as_tensor(
rot_mat,
dtype=ttnn.bfloat16,
layout=ttnn.TILE_LAYOUT,
layout=ttnn.ROW_MAJOR_LAYOUT,
mesh_mapper=ReplicateTensorToMesh(self.mesh_device),
)

Expand Down
Loading