[inference] add llama2 support (#4898)

* add llama2 support

* fix multi group bug

colossalai/inference/tensor_parallel/engine.py

@@ -1,7 +1,6 @@
 from typing import Any, Callable, List, Optional, Union
 
 import torch
-import torch.distributed as dist
 import torch.nn as nn
 from transformers import BloomForCausalLM, LlamaForCausalLM
 from transformers.generation import GenerationConfig
@@ -74,9 +73,14 @@ class TPInferEngine:
             model.config.num_hidden_layers if hasattr(model.config, "num_hidden_layers") else model.config.num_layers
         )
         self.layer_num = num_hidden_layers
-        self.multi_query_group_num = (
+
-            model.config.multi_query_group_num if hasattr(model.config, "multi_query_group_num") else 0
+        self.multi_query_group_num = 0
-        )
+
+        if hasattr(model.config, "multi_query_group_num"):
+            self.multi_query_group_num = model.config.multi_query_group_num
+
+        if hasattr(model.config, "num_key_value_heads"):
+            self.multi_query_group_num = model.config.num_key_value_heads
 
         self.tp_size = -1  # to be set with given shard config in self.prepare_shard_config
         self.cache_manager = None
@@ -97,6 +101,7 @@ class TPInferEngine:
         assert self.tp_size >= 1, "TP size not initialized without providing a valid ShardConfig"
         assert self.head_num % self.tp_size == 0, f"Cannot shard {self.head_num} heads with tp size {self.tp_size}"
         self.head_num //= self.tp_size  # update sharded number of heads
+
         if self.multi_query_group_num:
             # NOTE the logic of MQA tensor parallelism should be specified.
             assert (
@@ -116,13 +121,15 @@ class TPInferEngine:
 
     def _post_init_gptq_buffer(self, model: nn.Module) -> None:
         from colossalai.inference.quant.gptq.cai_gptq import CaiQuantLinear
+
         HAS_GPTQ_CUDA = False
         try:
             from colossalai.kernel.op_builder.gptq import GPTQBuilder
+
             gptq_cuda = GPTQBuilder().load()
             HAS_GPTQ_CUDA = True
         except ImportError:
-            warnings.warn('CUDA gptq is not installed')
+            warnings.warn("CUDA gptq is not installed")
             HAS_GPTQ_CUDA = False
 
         for name, submodule in model.named_modules():
@@ -130,8 +137,9 @@ class TPInferEngine:
                 self.max_dq_buffer_size = max(self.max_dq_buffer_size, submodule.qweight.numel() * 8)
 
                 if self.use_act_order:
-                    self.max_inner_outer_dim = max(self.max_inner_outer_dim, submodule.infeatures,
+                    self.max_inner_outer_dim = max(
-                                                   submodule.outfeatures)
+                        self.max_inner_outer_dim, submodule.infeatures, submodule.outfeatures
+                    )
                 self.bits = submodule.bits
         if not (HAS_GPTQ_CUDA and self.bits == 4):
             return
@@ -141,15 +149,16 @@ class TPInferEngine:
             max_input_len = self.max_input_len
         # The temp_state buffer is required to reorder X in the act-order case.
         # The temp_dq buffer is required to dequantize weights when using cuBLAS, typically for the prefill.
-        self.gptq_temp_state_buffer = torch.zeros((max_input_len, self.max_inner_outer_dim),
+        self.gptq_temp_state_buffer = torch.zeros(
-                                                  dtype=torch.float16,
+            (max_input_len, self.max_inner_outer_dim), dtype=torch.float16, device=torch.cuda.current_device()
-                                                  device=torch.cuda.current_device())
+        )
-        self.gptq_temp_dq_buffer = torch.zeros((1, self.max_dq_buffer_size),
+        self.gptq_temp_dq_buffer = torch.zeros(
-                                               dtype=torch.float16,
+            (1, self.max_dq_buffer_size), dtype=torch.float16, device=torch.cuda.current_device()
-                                               device=torch.cuda.current_device())
+        )
 
-        gptq_cuda.prepare_buffers(torch.device(torch.cuda.current_device()), self.gptq_temp_state_buffer,
+        gptq_cuda.prepare_buffers(
-                                  self.gptq_temp_dq_buffer)
+            torch.device(torch.cuda.current_device()), self.gptq_temp_state_buffer, self.gptq_temp_dq_buffer
+        )
         # Using the default from exllama repo here.
         matmul_recons_thd = 8
         matmul_fused_remap = False

colossalai/inference/tensor_parallel/modeling/_utils.py

@@ -45,7 +45,7 @@ def init_to_get_rotary(self, base=10000, use_elem=False):
     base = float(base)
 
     # NTK  ref: https://www.reddit.com/r/LocalLLaMA/comments/14lz7j5/ntkaware_scaled_rope_allows_llama_models_to_have/
-    ntk_alpha = float(os.environ.get("INFER_NTK_ALPHA", None))
+    ntk_alpha = os.environ.get("INFER_NTK_ALPHA", None)
 
     if ntk_alpha is not None:
         ntk_alpha = float(ntk_alpha)

colossalai/inference/tensor_parallel/modeling/llama.py

@@ -5,7 +5,13 @@ from transformers.modeling_outputs import BaseModelOutputWithPast
 from transformers.models.llama.modeling_llama import LlamaAttention, LlamaDecoderLayer, LlamaModel, LlamaRMSNorm
 
 from colossalai.inference.tensor_parallel.batch_infer_state import BatchInferState
-from colossalai.kernel.triton import llama_context_attn_fwd, rotary_embedding_fwd, token_attention_fwd
+from colossalai.kernel.triton import (
+    llama2_context_attn_fwd,
+    llama_context_attn_fwd,
+    rotary_embedding_fwd,
+    token_attention_fwd,
+)
+from colossalai.kernel.triton.token_attention_kernel import Llama2TokenAttentionForwards
 
 from ._utils import copy_kv_to_mem_cache
 
@@ -138,6 +144,7 @@ class LlamaInferenceForwards:
             seq_len = infer_state.seq_len
             infer_state.position_cos = torch.index_select(self._cos_cached, 0, seq_len - 1).view(seq_len.shape[0], -1)
             infer_state.position_sin = torch.index_select(self._sin_cached, 0, seq_len - 1).view(seq_len.shape[0], -1)
+            infer_state.other_kv_index = infer_state.block_loc[0, seq_length_with_past - 1].item()
 
         if inputs_embeds is None:
             inputs_embeds = self.embed_tokens(input_ids)
@@ -261,8 +268,8 @@ class LlamaInferenceForwards:
         # key_states_transposed [bs, num_heads, seq_len, head_dim/embed_size_per_head]
 
         query_states = self.q_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
-        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        key_states = self.k_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim)
-        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_heads, self.head_dim)
+        value_states = self.v_proj(hidden_states).view(bsz, q_len, self.num_key_value_heads, self.head_dim)
 
         # NOTE might want to revise
         #   need some way to record the length of past key values cache
@@ -274,11 +281,11 @@ class LlamaInferenceForwards:
         # print("shape ", cos.shape, query_states.view(-1, self.num_heads, self.head_dim).shape, )
 
         rotary_embedding_fwd(query_states.view(-1, self.num_heads, self.head_dim), cos, sin)
-        rotary_embedding_fwd(key_states.view(-1, self.num_heads, self.head_dim), cos, sin)
+        rotary_embedding_fwd(key_states.view(-1, self.num_key_value_heads, self.head_dim), cos, sin)
 
         query_states = query_states.reshape(-1, self.num_heads, self.head_dim)
-        key_states = key_states.reshape(-1, self.num_heads, self.head_dim)
+        key_states = key_states.reshape(-1, self.num_key_value_heads, self.head_dim)
-        value_states = value_states.reshape(-1, self.num_heads, self.head_dim)
+        value_states = value_states.reshape(-1, self.num_key_value_heads, self.head_dim)
 
         if infer_state.is_context_stage:
             # first token generation
@@ -294,6 +301,7 @@ class LlamaInferenceForwards:
 
             attn_output = torch.empty_like(query_states)
 
+            if self.num_key_value_groups == 1:
                 llama_context_attn_fwd(
                     query_states,
                     key_states,
@@ -303,6 +311,16 @@ class LlamaInferenceForwards:
                     infer_state.seq_len,
                     infer_state.cache_manager.past_key_values_length,
                 )
+            else:
+                llama2_context_attn_fwd(
+                    query_states,
+                    key_states,
+                    value_states,
+                    attn_output,
+                    infer_state.start_loc,
+                    infer_state.seq_len,
+                    infer_state.cache_manager.past_key_values_length,
+                )
         else:
             if infer_state.decode_is_contiguous:
                 # if decode is contiguous, then we copy to key cache and value cache in cache manager directly
@@ -330,6 +348,7 @@ class LlamaInferenceForwards:
             # (batch_size, seqlen, nheads, headdim)
             attn_output = torch.empty_like(query_states)
 
+            if self.num_key_value_groups == 1:
                 token_attention_fwd(
                     query_states,
                     infer_state.cache_manager.key_buffer[infer_state.decode_layer_id],
@@ -340,7 +359,18 @@ class LlamaInferenceForwards:
                     infer_state.seq_len,
                     infer_state.cache_manager.past_key_values_length,
                 )
-
+            else:
+                Llama2TokenAttentionForwards.token_attn(
+                    query_states,
+                    infer_state.cache_manager.key_buffer[infer_state.decode_layer_id],
+                    infer_state.cache_manager.value_buffer[infer_state.decode_layer_id],
+                    attn_output,
+                    infer_state.block_loc,
+                    infer_state.start_loc,
+                    infer_state.seq_len,
+                    infer_state.cache_manager.past_key_values_length,
+                    infer_state.other_kv_index,
+                )
         attn_output = attn_output.view(bsz, q_len, self.hidden_size)
 
         attn_output = self.o_proj(attn_output)

colossalai/kernel/triton/__init__.py

@@ -9,7 +9,7 @@ except ImportError:
 
 # There may exist import error even if we have triton installed.
 if HAS_TRITON:
-    from .context_attention import bloom_context_attn_fwd, llama_context_attn_fwd
+    from .context_attention import bloom_context_attn_fwd, llama2_context_attn_fwd, llama_context_attn_fwd
     from .copy_kv_cache_dest import copy_kv_cache_to_dest
     from .fused_layernorm import layer_norm
     from .gptq_triton import gptq_fused_linear_triton
@@ -20,6 +20,7 @@ if HAS_TRITON:
 
     __all__ = [
         "llama_context_attn_fwd",
+        "llama2_context_attn_fwd",
         "bloom_context_attn_fwd",
         "softmax",
         "layer_norm",

tests/test_infer/test_llama2_infer.py

@@ -0,0 +1,69 @@
+import os
+
+import pytest
+import torch
+from packaging import version
+from transformers import LlamaForCausalLM
+from transformers.models.llama.configuration_llama import LlamaConfig
+
+import colossalai
+from colossalai.inference.tensor_parallel.engine import TPInferEngine
+from colossalai.logging import disable_existing_loggers
+from colossalai.shardformer import ShardConfig
+from colossalai.testing import clear_cache_before_run, parameterize, rerun_if_address_is_in_use, spawn
+
+os.environ["TRANSFORMERS_NO_ADVISORY_WARNINGS"] = "true"
+TPSIZE = 2
+BATCH_SIZE = 8
+MAX_INPUT_LEN = 12
+MAX_OUTPUT_LEN = 100
+
+CUDA_SUPPORT = version.parse(torch.version.cuda) > version.parse("11.5")
+
+
+@parameterize(
+    "test_config",
+    [
+        {
+            "tp_size": TPSIZE,
+        }
+    ],
+)
+def run_llama_test(test_config):
+    llama_config = LlamaConfig(
+        num_hidden_layers=2, num_key_value_heads=8, bos_token_id=0, eos_token_id=1, vocab_size=1200, hidden_size=1024
+    )
+    model = LlamaForCausalLM(llama_config)
+    model = model.half()
+
+    shard_config = ShardConfig(
+        enable_tensor_parallelism=True if test_config["tp_size"] > 1 else False, inference_only=True
+    )
+    infer_engine = TPInferEngine(model, shard_config, BATCH_SIZE, MAX_INPUT_LEN, MAX_OUTPUT_LEN)
+    generate_kwargs = dict(max_new_tokens=MAX_OUTPUT_LEN, do_sample=False)
+
+    input_tokens = {
+        "input_ids": torch.randint(1, 1000, (BATCH_SIZE, MAX_INPUT_LEN), device="cuda"),
+        "attention_mask": torch.ones((BATCH_SIZE, MAX_INPUT_LEN), device="cuda"),
+    }
+    outputs = infer_engine.generate(input_tokens, **generate_kwargs)
+
+    assert outputs is not None
+
+
+def check_llama(rank, world_size, port):
+    disable_existing_loggers()
+    colossalai.launch(config={}, rank=rank, world_size=world_size, host="localhost", port=port, backend="nccl")
+    run_llama_test()
+
+
+@pytest.mark.skipif(not CUDA_SUPPORT, reason="kv-cache manager engine requires cuda version to be higher than 11.5")
+@pytest.mark.dist
+@rerun_if_address_is_in_use()
+@clear_cache_before_run()
+def test_llama():
+    spawn(check_llama, TPSIZE)
+
+
+if __name__ == "__main__":
+    test_llama()