Sync from v0.13

vllm/model_executor/models/jais.py

@@ -1,6 +1,8 @@
-# coding=utf-8
+# SPDX-License-Identifier: Apache-2.0
+# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
+
 # Adapted from
-# https://huggingface.co/core42/jais-30b-chat-v3/blob/main/modeling_jais.py
+# https://huggingface.co/inceptionai/jais-30b-chat-v3/blob/main/modeling_jais.py
 # Copyright 2023 The vLLM team.
 # Copyright 2023 the Jais authors and HuggingFace Inc. team.  All rights
 # reserved.
@@ -20,62 +22,77 @@
 """Inference-only Jais model compatible with HuggingFace weights."""
 
 import math
-from typing import Iterable, List, Optional, Tuple
+from collections.abc import Iterable
+from itertools import islice
 
 import torch
 from torch import nn
 
-from vllm.attention import Attention, AttentionMetadata
-from vllm.distributed import (get_tensor_model_parallel_rank,
-                              get_tensor_model_parallel_world_size)
-from vllm.model_executor.layers.linear import (ColumnParallelLinear,
-                                               QKVParallelLinear,
-                                               RowParallelLinear)
+from vllm.attention.layer import Attention
+from vllm.compilation.decorators import support_torch_compile
+from vllm.config import CacheConfig, VllmConfig
+from vllm.distributed import (
+    get_pp_group,
+    get_tensor_model_parallel_rank,
+    get_tensor_model_parallel_world_size,
+)
+from vllm.model_executor.layers.linear import (
+    ColumnParallelLinear,
+    QKVParallelLinear,
+    RowParallelLinear,
+)
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
-from vllm.model_executor.layers.quantization.base_config import (
-    QuantizationConfig)
-from vllm.model_executor.layers.sampler import Sampler
+from vllm.model_executor.layers.quantization import QuantizationConfig
 from vllm.model_executor.layers.vocab_parallel_embedding import (
-    VocabParallelEmbedding)
+    ParallelLMHead,
+    VocabParallelEmbedding,
+)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.sampling_metadata import SamplingMetadata
-from vllm.sequence import SamplerOutput
+from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import JAISConfig
 
+from .interfaces import SupportsPP
+from .utils import (
+    is_pp_missing_parameter,
+    make_empty_intermediate_tensors_factory,
+    make_layers,
+    maybe_prefix,
+)
+
 
 class SwiGLUActivation(nn.Module):
-
     def forward(self, x1: torch.Tensor, x2: torch.Tensor) -> torch.Tensor:
         return x1 * nn.functional.silu(x2)
 
 
 def _get_alibi_slopes(n):
-
     def get_slopes_power_of_2(n):
-        start = 2**(-(2**-(math.log2(n) - 3)))
+        start = 2 ** (-(2 ** -(math.log2(n) - 3)))
         ratio = start
         return [start * ratio**i for i in range(n)]
 
     if math.log2(n).is_integer():
         return get_slopes_power_of_2(n)
     else:
-        closest_power_of_2 = 2**math.floor(math.log2(n))
-        return (get_slopes_power_of_2(closest_power_of_2) + _get_alibi_slopes(
-            2 * closest_power_of_2)[0::2][:n - closest_power_of_2])
+        closest_power_of_2 = 2 ** math.floor(math.log2(n))
+        return (
+            get_slopes_power_of_2(closest_power_of_2)
+            + _get_alibi_slopes(2 * closest_power_of_2)[0::2][: n - closest_power_of_2]
+        )
 
 
 class JAISAttention(nn.Module):
-
     def __init__(
         self,
         config: JAISConfig,
-        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ):
         super().__init__()
         self.hidden_size = config.hidden_size
         total_num_heads = config.num_attention_heads
-        tensor_model_parallel_world_size = (
-            get_tensor_model_parallel_world_size())
+        tensor_model_parallel_world_size = get_tensor_model_parallel_world_size()
         assert total_num_heads % tensor_model_parallel_world_size == 0
         self.num_heads = total_num_heads // tensor_model_parallel_world_size
         self.head_dim = self.hidden_size // total_num_heads
@@ -90,12 +107,14 @@ class JAISAttention(nn.Module):
             total_num_heads,
             bias=True,
             quant_config=quant_config,
+            prefix=f"{prefix}.c_attn",
         )
         self.c_proj = RowParallelLinear(
             self.hidden_size,
             self.hidden_size,
             bias=True,
             quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
         )
 
         tp_rank = get_tensor_model_parallel_rank()
@@ -108,28 +127,29 @@ class JAISAttention(nn.Module):
             self.head_dim,
             scale=self.scale,
             alibi_slopes=alibi_slopes,
+            cache_config=cache_config,
+            quant_config=quant_config,
+            prefix=f"{prefix}.attn",
         )
 
     def forward(
         self,
         hidden_states: torch.Tensor,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         qkv, _ = self.c_attn(hidden_states)
         q, k, v = qkv.chunk(chunks=3, dim=-1)
-        attn_output = self.attn(q, k, v, kv_cache, attn_metadata)
+        attn_output = self.attn(q, k, v)
         attn_output, _ = self.c_proj(attn_output)
         return attn_output
 
 
 class JAISMLP(nn.Module):
-
     def __init__(
         self,
         intermediate_size: int,
         config: JAISConfig,
-        quant_config: Optional[QuantizationConfig] = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ):
         super().__init__()
         hidden_size = config.hidden_size
@@ -139,18 +159,25 @@ class JAISMLP(nn.Module):
             intermediate_size,
             bias=True,
             quant_config=quant_config,
+            prefix=f"{prefix}.c_fc",
+        )
+        self.c_fc2 = (
+            ColumnParallelLinear(
+                hidden_size,
+                intermediate_size,
+                bias=True,
+                quant_config=quant_config,
+                prefix=f"{prefix}.c_fc2",
+            )
+            if self.swiglu
+            else None
         )
-        self.c_fc2 = (ColumnParallelLinear(
-            hidden_size,
-            intermediate_size,
-            bias=True,
-            quant_config=quant_config,
-        ) if self.swiglu else None)
         self.c_proj = RowParallelLinear(
             intermediate_size,
             hidden_size,
             bias=True,
             quant_config=quant_config,
+            prefix=f"{prefix}.c_proj",
         )
 
         self.act = SwiGLUActivation()
@@ -159,41 +186,42 @@ class JAISMLP(nn.Module):
         if self.swiglu:
             hidden_states2, _ = self.c_fc2(hidden_states)
         hidden_states, _ = self.c_fc(hidden_states)
-        hidden_states = (self.act(hidden_states, hidden_states2)
-                         if self.swiglu else self.act(hidden_states))
+        hidden_states = (
+            self.act(hidden_states, hidden_states2)
+            if self.swiglu
+            else self.act(hidden_states)
+        )
         hidden_states, _ = self.c_proj(hidden_states)
         return hidden_states
 
 
 class JAISBlock(nn.Module):
-
     def __init__(
         self,
         config: JAISConfig,
-        quant_config: Optional[QuantizationConfig] = None,
+        cache_config: CacheConfig | None = None,
+        quant_config: QuantizationConfig | None = None,
+        prefix: str = "",
     ):
         super().__init__()
         hidden_size = config.hidden_size
-        inner_dim = (config.n_inner if config.n_inner is not None else 4 *
-                     hidden_size)
+        inner_dim = config.n_inner if config.n_inner is not None else 4 * hidden_size
 
         self.ln_1 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
-        self.attn = JAISAttention(config, quant_config)
+        self.attn = JAISAttention(
+            config, cache_config, quant_config, prefix=f"{prefix}.attn"
+        )
         self.ln_2 = nn.LayerNorm(hidden_size, eps=config.layer_norm_epsilon)
-        self.mlp = JAISMLP(inner_dim, config, quant_config)
+        self.mlp = JAISMLP(inner_dim, config, quant_config, prefix=f"{prefix}.mlp")
 
     def forward(
         self,
         hidden_states: torch.Tensor,
-        kv_cache: torch.Tensor,
-        attn_metadata: AttentionMetadata,
     ) -> torch.Tensor:
         residual = hidden_states
         hidden_states = self.ln_1(hidden_states)
         attn_output = self.attn(
             hidden_states=hidden_states,
-            kv_cache=kv_cache,
-            attn_metadata=attn_metadata,
         )
         # residual connection
         hidden_states = attn_output + residual
@@ -206,105 +234,136 @@ class JAISBlock(nn.Module):
         return hidden_states
 
 
+@support_torch_compile
 class JAISModel(nn.Module):
-
-    def __init__(
-        self,
-        config: JAISConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-    ):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
+
+        config = vllm_config.model_config.hf_config
+        cache_config = vllm_config.cache_config
+        quant_config = vllm_config.quant_config
+
         self.config = config
         assert not config.add_cross_attention
         assert not config.scale_attn_by_inverse_layer_idx
         assert not config.reorder_and_upcast_attn
         self.embed_dim = config.hidden_size
         self.wte = VocabParallelEmbedding(config.vocab_size, self.embed_dim)
-        self.wpe = (nn.Embedding(config.max_position_embeddings,
-                                 self.embed_dim)
-                    if config.position_embedding_type != "alibi" else None)
+        self.wpe = (
+            nn.Embedding(config.max_position_embeddings, self.embed_dim)
+            if config.position_embedding_type != "alibi"
+            else None
+        )
         if hasattr(config, "embeddings_scale"):
             self.embeddings_scale = config.embeddings_scale
         else:
             self.embeddings_scale = config.mup_embeddings_scale
-        self.h = nn.ModuleList([
-            JAISBlock(config, quant_config)
-            for _ in range(config.num_hidden_layers)
-        ])
+
+        self.start_layer, self.end_layer, self.h = make_layers(
+            config.num_hidden_layers,
+            lambda prefix: JAISBlock(
+                config=config,
+                cache_config=cache_config,
+                quant_config=quant_config,
+                prefix=prefix,
+            ),
+            prefix=f"{prefix}.h",
+        )
+
         self.ln_f = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_epsilon)
+        self.make_empty_intermediate_tensors = make_empty_intermediate_tensors_factory(
+            ["hidden_states"], config.n_embd
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.wte(input_ids)
 
     def forward(
         self,
         input_ids: torch.Tensor,
         position_ids: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-    ) -> torch.Tensor:
-        inputs_embeds = self.wte(input_ids)
-        if self.wpe is not None:
-            position_embeds = self.wpe(position_ids)
-            hidden_states = inputs_embeds + position_embeds
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> IntermediateTensors | torch.Tensor:
+        if get_pp_group().is_first_rank:
+            if inputs_embeds is None:
+                inputs_embeds = self.embed_input_ids(input_ids)
+            if self.wpe is not None:
+                position_embeds = self.wpe(position_ids)
+                hidden_states = inputs_embeds + position_embeds
+            else:
+                hidden_states = inputs_embeds
+            hidden_states *= torch.tensor(
+                float(self.embeddings_scale), dtype=hidden_states.dtype
+            )
         else:
-            hidden_states = inputs_embeds
-        hidden_states *= torch.tensor(float(self.embeddings_scale),
-                                      dtype=hidden_states.dtype)
+            assert intermediate_tensors is not None
+            hidden_states = intermediate_tensors["hidden_states"]
 
-        for i in range(len(self.h)):
-            layer = self.h[i]
-            hidden_states = layer(hidden_states, kv_caches[i], attn_metadata)
+        for layer in islice(self.h, self.start_layer, self.end_layer):
+            hidden_states = layer(hidden_states)
+
+        if not get_pp_group().is_last_rank:
+            return IntermediateTensors({"hidden_states": hidden_states})
 
         hidden_states = self.ln_f(hidden_states)
         return hidden_states
 
 
-class JAISLMHeadModel(nn.Module):
-
-    def __init__(
-        self,
-        config: JAISConfig,
-        quant_config: Optional[QuantizationConfig] = None,
-    ):
+class JAISLMHeadModel(nn.Module, SupportsPP):
+    def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
         super().__init__()
+        config = vllm_config.model_config.hf_config
+        quant_config = vllm_config.quant_config
         self.config = config
         self.quant_config = quant_config
-        self.transformer = JAISModel(config, quant_config)
-        self.lm_head_weight = self.transformer.wte.weight
+        self.transformer = JAISModel(
+            vllm_config=vllm_config, prefix=maybe_prefix(prefix, "transformer")
+        )
+        if self.config.tie_word_embeddings:
+            self.lm_head = self.transformer.wte
+        else:
+            self.lm_head = ParallelLMHead(
+                self.config.vocab_size,
+                self.config.hidden_size,
+                prefix=maybe_prefix(prefix, "lm_head"),
+            )
         if hasattr(config, "width_scale"):
             self.output_logits_scale = config.width_scale
         else:
-            self.output_logits_scale = (config.mup_output_alpha *
-                                        config.mup_width_scale)
-        self.logits_processor = LogitsProcessor(vocab_size=config.vocab_size,
-                                                scale=self.output_logits_scale)
-        self.sampler = Sampler()
+            self.output_logits_scale = config.mup_output_alpha * config.mup_width_scale
+        self.logits_processor = LogitsProcessor(
+            vocab_size=config.vocab_size, scale=self.output_logits_scale
+        )
+        self.make_empty_intermediate_tensors = (
+            self.transformer.make_empty_intermediate_tensors
+        )
+
+    def embed_input_ids(self, input_ids: torch.Tensor) -> torch.Tensor:
+        return self.transformer.embed_input_ids(input_ids)
 
     def forward(
         self,
         input_ids: torch.Tensor,
         positions: torch.Tensor,
-        kv_caches: List[torch.Tensor],
-        attn_metadata: AttentionMetadata,
-    ) -> torch.Tensor:
-        hidden_states = self.transformer(input_ids, positions, kv_caches,
-                                         attn_metadata)
+        intermediate_tensors: IntermediateTensors | None = None,
+        inputs_embeds: torch.Tensor | None = None,
+    ) -> IntermediateTensors | torch.Tensor:
+        hidden_states = self.transformer(
+            input_ids, positions, intermediate_tensors, inputs_embeds
+        )
         return hidden_states
 
-    def compute_logits(self, hidden_states: torch.Tensor,
-                       sampling_metadata: SamplingMetadata) -> torch.Tensor:
-        logits = self.logits_processor(self.lm_head_weight, hidden_states,
-                                       sampling_metadata)
+    def compute_logits(
+        self,
+        hidden_states: torch.Tensor,
+    ) -> torch.Tensor | None:
+        logits = self.logits_processor(self.lm_head, hidden_states)
         return logits
 
-    def sample(
-        self,
-        logits: torch.Tensor,
-        sampling_metadata: SamplingMetadata,
-    ) -> Optional[SamplerOutput]:
-        next_tokens = self.sampler(logits, sampling_metadata)
-        return next_tokens
-
-    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+    def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
         params_dict = dict(self.named_parameters(remove_duplicate=False))
+        loaded_params: set[str] = set()
         for name, loaded_weight in weights:
             if "lm_head.weight" in name:
                 # GPT-2 ties the weights of the embedding layer and the final
@@ -318,6 +377,10 @@ class JAISLMHeadModel(nn.Module):
                 continue
             if not name.startswith("transformer."):
                 name = "transformer." + name
+
+            if is_pp_missing_parameter(name, self):
+                continue
+
             param = params_dict[name]
             # The HF's GPT-2 implementation uses Conv1D instead of Linear.
             # Because of this, we need to transpose the weights.
@@ -328,6 +391,7 @@ class JAISLMHeadModel(nn.Module):
                 if not name.endswith(".weight"):
                     continue
                 loaded_weight = loaded_weight.t()
-            weight_loader = getattr(param, "weight_loader",
-                                    default_weight_loader)
+            weight_loader = getattr(param, "weight_loader", default_weight_loader)
             weight_loader(param, loaded_weight)
+            loaded_params.add(name)
+        return loaded_params