[Llama3-text vLLM integration] Modify Llama3 text model (new and old …

…codebase) forward apis for vLLM compatibility (#16292)

models/demos/llama3/tt/generator.py

@@ -55,6 +55,7 @@ def prefill_forward_text(self, tokens: torch.Tensor, page_table=None, kv_cache=N
             ), "page_table must be a torch.Tensor when passing into prefill_forward"
 
         for user_id in range(batch):
+            logger.info(f"Prefilling User {user_id + 1}")
             seq_len = prompt_lens[user_id]
             last_token_idx = seq_len - 1
 
@@ -76,6 +77,8 @@ def prefill_forward_text(self, tokens: torch.Tensor, page_table=None, kv_cache=N
             # Since we give unpadded_seq_len, only the tile containing the last token is returned
             output_logits[user_id] = logits
 
+        logger.info(f"Finished prefill for all users up to {batch_seq_len} tokens, Starting decode...")
+
         return output_logits
 
     def prefill_forward_single_user_text(self, tokens, page_table, user_id, last_token_idx, kv_cache=None):
@@ -162,14 +165,40 @@ def prefill_forward_single_user_text(self, tokens, page_table, user_id, last_tok
             return logits
 
     def decode_forward_text(
+        self,
+        tokens,
+        start_pos,
+        page_table=None,
+        kv_cache=None,
+        enable_trace=True,
+        read_from_device=True,
+    ):
+        decode_kwargs = {
+            "current_pos": start_pos,
+            "tokens": tokens,
+            "page_table": page_table,
+            "kv_cache": kv_cache,
+        }
+        if enable_trace:
+            tt_logits = self._easy_trace_text(**decode_kwargs)
+        else:
+            tt_logits = self._decode_forward_no_trace_text(**decode_kwargs)
+
+        if read_from_device:
+            return self.read_decode_output(tt_logits, tokens.shape[0])
+        else:
+            return tt_logits
+
+    def _decode_forward_no_trace_text(
         self,
         tokens,
         current_pos,
         page_table=None,
+        kv_cache=None,
     ):
         """
         Performs text decode step.
-        Returns logits
+        Returns tt_logits on device
         """
         tt_tokens, tt_current_pos, tt_rot_mats, tt_page_table = self.model.prepare_inputs_decode(
             tokens, current_pos, page_table
@@ -180,38 +209,41 @@ def decode_forward_text(
             tt_current_pos,
             rot_mats=tt_rot_mats,
             page_table=tt_page_table,
+            kv_cache=kv_cache,
         )
 
-        logits = self.model.process_output_decode(tt_logits)
-        return logits
+        return tt_logits
 
-    def capture_trace_text(
+    def _capture_trace_text(
         self,
         tokens,
         current_pos,
         page_table=None,
+        kv_cache=None,
     ):
         """
         Captures a trace for the decode_forward method.
         """
 
         # Compile run
-        self.decode_forward_text(tokens, current_pos, page_table)
+        self._decode_forward_no_trace_text(tokens, current_pos, page_table=page_table, kv_cache=kv_cache)
+        logger.info("Done Compiling Model")
 
         # Get inputs ready for trace run
-        host_inputs = self.model.prepare_decode_inputs_host(tokens, current_pos, page_table)
+        host_inputs = self.model.prepare_decode_inputs_host(tokens, current_pos, page_table=page_table)
 
         device_inputs = copy_host_to_device(host_inputs, mesh_device=self.mesh_device)
 
         trace_id = ttnn.begin_trace_capture(self.mesh_device, cq_id=0)
         transformed_inputs = self.model.transform_decode_inputs_device(*device_inputs)
-        tt_out_trace = self.model.ttnn_decode_forward(*transformed_inputs)
+        tt_out_trace = self.model.ttnn_decode_forward(*transformed_inputs, kv_cache=kv_cache)
 
         ttnn.end_trace_capture(self.mesh_device, trace_id, cq_id=0)
+        logger.info("Done Capturing Decode Trace")
 
         return trace_id, tt_out_trace, *device_inputs
 
-    def decode_forward_trace_text(
+    def _decode_forward_trace_text(
         self,
         trace_id,
         device_inputs,
@@ -220,6 +252,9 @@ def decode_forward_trace_text(
         current_pos,
         page_table=None,
     ):
+        """
+        Executes the trace for the decode_forward method but does not read back outputs.
+        """
         host_inputs = self.model.prepare_decode_inputs_host(tokens, current_pos, page_table)
 
         device_inputs = copy_host_to_device(
@@ -229,9 +264,36 @@ def decode_forward_trace_text(
 
         ttnn.execute_trace(self.mesh_device, trace_id, cq_id=0, blocking=False)
 
-        logits = self.model.process_output_decode(tt_out_trace)
+        return tt_out_trace
 
-        return logits
+    def _easy_trace_text(
+        self,
+        tokens,
+        current_pos,
+        page_table=None,
+        kv_cache=None,
+    ):
+        """
+        Tracing is easy! Just call this method and we'll handle tracing for you.
+        """
+        if not hasattr(self, "trace_id_text"):
+            trace_id, tt_out_trace, *device_inputs = self._capture_trace_text(
+                tokens, current_pos, page_table=page_table, kv_cache=kv_cache
+            )
+            self.trace_id_text = trace_id
+            self.trace_inputs_text = device_inputs
+            self.trace_output_text = tt_out_trace
+
+        trace_logits_rm = self._decode_forward_trace_text(
+            self.trace_id_text,
+            self.trace_inputs_text,
+            self.trace_output_text,
+            tokens,
+            current_pos,
+            page_table=page_table,
+        )
+
+        return trace_logits_rm
 
     def _prefill_forward_single_user(
         self,
@@ -325,7 +387,7 @@ def prefill_forward(
         output_full_text_row_masked_out_masks = []
 
         for user_id in range(batch):
-            print(f"Prefilling User {user_id}")
+            logger.info(f"Prefilling User {user_id + 1}")
             seq_len = prompt_lens[user_id]
             (
                 xattn_caches,
@@ -833,3 +895,15 @@ def _get_prefill_user_page_table(self, page_table, kv_cache, prefill_len):
         block_size = get_block_size(kv_cache)
         num_blocks = num_blocks_in_seq(prefill_len, block_size)
         return page_table[:, :num_blocks]
+
+    ## Destructor (used to delete ttnn trace if exists)
+
+    def __del__(self):
+        if hasattr(self, "trace_id"):
+            ttnn.release_trace(self.mesh_device, self.trace_id)
+
+        if hasattr(self, "trace_id_text"):
+            ttnn.release_trace(self.mesh_device, self.trace_id_text)
+
+        if hasattr(super(LlamaGenerator, self), "__del__"):
+            super().__del__()

models/demos/llama3/tt/generator_vllm.py

@@ -6,8 +6,12 @@
 import torch
 import PIL
 from llama_models.llama3.api.chat_format import create_vision_mask
+from llama_models.llama3.api.tokenizer import Tokenizer
+import ttnn
 
 from models.demos.llama3.tt.generator import LlamaGenerator
+from models.demos.llama3.tt.llama_model import TtTransformer
+from models.demos.llama3.tt.model_config import LlamaOptimizations, TtModelArgs
 from models.demos.llama3.demo.simple_vision_demo import create_multimodal_model
 from models.utility_functions import nearest_32
 
@@ -106,3 +110,47 @@ def prefill_forward(
             kv_cache=kv_cache,
             cross_page_table=cross_page_table,
         )
+
+
+class TtLlamaForCausalLM(LlamaGenerator):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    @classmethod
+    def initialize_vllm_model(cls, hf_config, mesh_device, max_batch_size, n_layers=None):
+        instruct_mode = "Instruct" in hf_config._name_or_path
+        max_seq_len = 131072  # TODO: modify this for different models/devices
+        optimizations = LlamaOptimizations.performance  # TODO: maybe change to accuracy
+        dtype = ttnn.bfloat8_b
+
+        # Load model args, weights
+        model_args = TtModelArgs(
+            mesh_device,
+            instruct=instruct_mode,
+            max_batch_size=max_batch_size,
+            optimizations=optimizations,
+            max_seq_len=max_seq_len,
+        )
+        if n_layers is not None:
+            model_args.n_layers = n_layers
+        state_dict = model_args.load_state_dict()
+
+        tt_model = TtTransformer(
+            args=model_args,
+            mesh_device=mesh_device,
+            dtype=dtype,
+            state_dict=state_dict,
+            weight_cache_path=model_args.weight_cache_path(dtype),
+            use_paged_kv_cache=True,
+        )
+        return cls(tt_model, model_args, mesh_device)
+
+    @property
+    def cache_path(self):
+        return self.model_args.model_cache_path
+
+    def prefill_forward(self, *args, **kwargs):
+        return super().prefill_forward_text(*args, **kwargs)
+
+    def decode_forward(self, *args, **kwargs):
+        return super().decode_forward_text(*args, **kwargs)

models/demos/llama3/tt/llama_common.py

@@ -322,12 +322,12 @@ def sample_host(tt_input, mesh_device, temperature=0.6, top_p=0.08, on_host=True
 
 def get_padded_prefill_len(seq_len):
     """
-    If seq_len is less than 32, pad to 32
-    If seq_len is more than 32, pad to whichever is smaller: a power of 2 or a multiple of 1024
+    If seq_len is less than 128, pad to 128
+    If seq_len is more than 128, pad to whichever is smaller: a power of 2 or a multiple of 1024
     TODO: Generalize for max_mm_seq_len different from 1024
     """
-    if seq_len <= 32:
-        return 32
+    if seq_len <= 128:
+        return 128
     pow_2_pad = nearest_pow_2(seq_len)
     mult_1024_pad = 1024 * math.ceil(seq_len / 1024)
     min_extended_pad = min(pow_2_pad, mult_1024_pad)

models/demos/llama3/tt/llama_model.py

@@ -7,6 +7,7 @@
 import ttnn
 import torch
 import torch.nn as nn
+from tqdm import tqdm
 from models.demos.llama3.tt.llama_decoder import TtTransformerBlock
 from models.common.rmsnorm import RMSNorm
 import ttnn
@@ -28,6 +29,7 @@ def __init__(
         state_dict,
         weight_cache_path,
         paged_attention_config=None,
+        use_paged_kv_cache=False,
     ):
         super().__init__()
         self.args = args
@@ -69,8 +71,9 @@ def __init__(
                 layer_num=i,
                 transformation_mats=self.trans_mats_dict,
                 paged_attention_config=paged_attention_config,
+                use_paged_kv_cache=use_paged_kv_cache,
             )
-            for i in range(self.n_layers)
+            for i in tqdm(range(self.n_layers))
         ]
         self.norm = DistributedNorm(
             RMSNorm(
@@ -107,9 +110,8 @@ def prepare_inputs_prefill(self, tokens, start_pos=0, page_table=None, chunk_pag
 
         tokens = tokens.reshape(1, 1, 1, -1)
         S = tokens.shape[-1]
-        dims = (None, -1) if self.args.is_galaxy else (None, None)
+        dims = (None, None)  # replicate
         mesh_mapper = ttnn.ShardTensor2dMesh(self.mesh_device, dims=dims, mesh_shape=self.args.cluster_shape)
-
         tokens = ttnn.from_torch(
             tokens,
             device=self.mesh_device,
@@ -170,19 +172,20 @@ def prepare_decode_inputs_host(self, tokens, current_pos, page_table=None):
         Inputs are torch tensors or python types. Outputs are ttnn tensors on host.
         NOTE: Tokens and current_pos are padded to batch
         """
-        B = tokens.shape[-1]
+        B = tokens.shape[0]
         assert current_pos.shape[0] == B, "Batch size mismatch"
         assert B == self.args.max_batch_size, "Batch size must be equal to max_batch_size"
 
         dims = (None, -1) if self.args.is_galaxy else (None, None)
         mesh_mapper = ttnn.ShardTensor2dMesh(self.mesh_device, dims=dims, mesh_shape=self.args.cluster_shape)
 
         tokens = ttnn.from_torch(
-            tokens,
+            tokens.view(-1),
             device=None,
             dtype=ttnn.uint32,
             mesh_mapper=mesh_mapper,
         )
+        tokens = ttnn.unsqueeze_to_4D(tokens)
 
         rot_current_pos = torch.maximum(
             current_pos, torch.tensor(0, dtype=torch.int64)
@@ -240,17 +243,19 @@ def process_output_prefill(self, tt_out, last_token_idx):
         )[0, 0, last_token_idx, :]
         return logits
 
-    def process_output_decode(self, tt_out):
+    def process_output_decode(self, tt_out, B, S=1):
         """
         Input is ttnn device tensor of logits. Output is torch logits tensor
         """
         if self.args.num_devices > 1:
             tt_out = ttnn.all_gather(tt_out, dim=3, num_links=1, topology=ttnn.Topology.Linear)
-        tt_out_rm = ttnn.untilize(tt_out, use_multicore=True)
+        tt_out = ttnn.untilize(tt_out, use_multicore=True)
         if self.args.num_devices > 1:
-            return ttnn.to_torch(ttnn.get_device_tensors(tt_out_rm)[0]).float()
+            tt_out = ttnn.to_torch(ttnn.get_device_tensors(tt_out)[0]).float()
         else:
-            return ttnn.to_torch(tt_out_rm).float()
+            tt_out = ttnn.to_torch(tt_out).float()
+        tt_out = tt_out[:, :, :B, :].view(B, S, -1)
+        return tt_out
 
     def ttnn_prefill_forward(
         self,
@@ -280,7 +285,14 @@ def ttnn_prefill_forward(
             kv_cache=kv_cache,
         )
 
-    def ttnn_decode_forward(self, x, current_pos, rot_mats, page_table=None, kv_cache=None):
+    def ttnn_decode_forward(
+        self,
+        x,
+        current_pos,
+        rot_mats,
+        page_table=None,
+        kv_cache=None,
+    ):
         """
         This method will take device tensors and any other args to run forward.
         It returns ttnn device tensors.

models/demos/llama3/tt/llama_rope.py

@@ -132,7 +132,7 @@ def get_rot_idxs(self, position_idxs, on_host=False):
                 position_idxs,
                 dtype=ttnn.uint32,
                 layout=ttnn.ROW_MAJOR_LAYOUT,
-                mesh_mapper=ReplicateTensorToMesh(self.device) if self.num_devices > 1 else None,
+                mesh_mapper=ReplicateTensorToMesh(self.device) if self.is_mesh_device else None,
             )
         else:  # On device
             rot_idxs = ttnn.as_tensor(

models/demos/t3000/llama2_70b/tt/generator_vllm.py

@@ -2,10 +2,10 @@
 
 # SPDX-License-Identifier: Apache-2.0
 
-from typing import Union
 from dataclasses import dataclass
 from pathlib import Path
 import json
+import torch
 
 from models.demos.t3000.llama2_70b.tt.llama_generation import TtLlamaModelForGeneration
 from models.demos.t3000.llama2_70b.tt.llama_common import (
@@ -41,6 +41,10 @@ class TTArgs:
         model_config, ckpt_dir, _, cache_path = setup_llama_env(
             llama_version=llama_version,
         )
+
+        mesh_rows = t3k_mesh_device.shape.num_rows
+        mesh_cols = t3k_mesh_device.shape.num_cols
+        assert mesh_rows == 2 and mesh_cols == 4, f"Invalid mesh device shape: {mesh_rows}x{mesh_cols}"
         check_mesh_device(t3k_mesh_device, model_config)
 
         # initialize arg classes
@@ -67,3 +71,6 @@ class TTArgs:
     @property
     def cache_path(self):
         return self.tt_model.cache_path
+
+    def prefill_forward(self, tokens: torch.Tensor, page_table, kv_cache, prompt_lens):
+        return super().prefill_forward(tokens, 0, page_table, kv_cache, prompt_lens)