|
| 1 | +#!/usr/bin/env python |
| 2 | +# Copyright (c) 2025 Oracle and/or its affiliates. |
| 3 | +# Licensed under the Universal Permissive License v 1.0 as shown at https://oss.oracle.com/licenses/upl/ |
| 4 | +from typing import Optional |
| 5 | + |
| 6 | +from pydantic import BaseModel, Field |
| 7 | + |
| 8 | +from ads.aqua.app import logger |
| 9 | +from ads.aqua.shaperecommend.constants import LLAMA_REQUIRED_FIELDS, MOE_REQUIRED_FIELDS |
| 10 | +from ads.aqua.shaperecommend.llm_config import LLMConfig |
| 11 | + |
| 12 | + |
| 13 | +class MemoryEstimator(BaseModel): |
| 14 | + """ |
| 15 | + The generic estimator for Transformer Architecture models (OPT/ Bloom) |
| 16 | + Used as a fallback estimator if model identified is not a MoE or GQA Architecture Model. |
| 17 | + Has properties to estimate the KV Cache size, Model size, and total footprint (KV Cache + Model size) |
| 18 | + """ |
| 19 | + |
| 20 | + llm_config: LLMConfig = Field( |
| 21 | + ..., |
| 22 | + description="The model's config.json file with the necessary parameters for model size and KV cache estimation/", |
| 23 | + ) |
| 24 | + batch_size: int = ( |
| 25 | + 1 # we assume that estimation for batch sizes are not supported yet |
| 26 | + ) |
| 27 | + seq_len: Optional[int] = Field( |
| 28 | + 4096, description="The max-seq-len to estimate the size of the KV cache." |
| 29 | + ) |
| 30 | + |
| 31 | + @property |
| 32 | + def kv_cache_memory(self) -> float: |
| 33 | + """ |
| 34 | + Estimates the KV cache size (in GB) using the LLM config.json parameters. |
| 35 | +
|
| 36 | + Uses num_attention_heads (assumes no GQA, each attention head has its own query, key, value) for estimation |
| 37 | + """ |
| 38 | + seq_len = self.seq_len or self.llm_config.max_seq_len |
| 39 | + c = self.llm_config |
| 40 | + kv_cache_dtype_bytes = ( |
| 41 | + c.bytes_per_parameter |
| 42 | + ) # vLLM uses model's weight/quantization applied to KV cache |
| 43 | + |
| 44 | + total_bytes = ( |
| 45 | + self.batch_size |
| 46 | + * c.num_hidden_layers |
| 47 | + * 2 |
| 48 | + * c.num_attention_heads |
| 49 | + * seq_len |
| 50 | + * c.head_dim |
| 51 | + * kv_cache_dtype_bytes |
| 52 | + ) |
| 53 | + return total_bytes / 1e9 |
| 54 | + |
| 55 | + @property |
| 56 | + def model_memory(self) -> float: |
| 57 | + """ |
| 58 | + Estimates the model size (in GB) based on estimating the model parameter size and model weights |
| 59 | +
|
| 60 | + Model Parameter estimation: Standard decoder-only, untied/tied embeddings possible |
| 61 | + """ |
| 62 | + c = self.llm_config |
| 63 | + embedding_count = 1 if getattr(c, "tie_word_embeddings", True) else 2 |
| 64 | + embedding_params = ( |
| 65 | + embedding_count * c.vocab_size * c.hidden_size |
| 66 | + ) # input and output untied |
| 67 | + layer_params = 12 * c.num_hidden_layers * (c.hidden_size**2) # GPT-style |
| 68 | + num_params = layer_params + embedding_params |
| 69 | + |
| 70 | + return num_params * c.bytes_per_parameter / 1e9 |
| 71 | + |
| 72 | + # @property |
| 73 | + # def model_overhead(self) -> float: |
| 74 | + # overhead = max(1, math.ceil(0.0 * self.model_memory)) |
| 75 | + # return overhead |
| 76 | + |
| 77 | + @property |
| 78 | + def total_memory(self) -> float: |
| 79 | + """ |
| 80 | + Computes the total memory footprint of the model (KV cache & model size from estimated parameters) |
| 81 | + """ |
| 82 | + return self.model_memory + self.kv_cache_memory |
| 83 | + |
| 84 | + |
| 85 | +# Specialized estimators: |
| 86 | +class LlamaMemoryEstimator(MemoryEstimator): |
| 87 | + """ |
| 88 | + Estimator for GQA-type architectures. Handles tied (memory savings) and untied embeddings, |
| 89 | + and uses grouped attention (GQA) for more efficient KV cache memory estimation. |
| 90 | +
|
| 91 | + KV cache: Use num_attention_heads (assumes GQA) |
| 92 | + Model Parameter estimation: Standard decoder-only, untied/tied embeddings possible |
| 93 | + """ |
| 94 | + |
| 95 | + @property |
| 96 | + def model_memory(self) -> float: |
| 97 | + """ |
| 98 | + Returns estimated model parameter memory (in GB), accurately accounting |
| 99 | + for Llama-style attention and MLP, and tied or untied embeddings. |
| 100 | + """ |
| 101 | + c = self.llm_config |
| 102 | + |
| 103 | + embedding_params, attn_params = self._calc_attn_embed_params() |
| 104 | + |
| 105 | + # MLP params |
| 106 | + gate_proj = c.hidden_size * c.intermediate_size |
| 107 | + up_proj = c.hidden_size * c.intermediate_size |
| 108 | + down_proj = c.intermediate_size * c.hidden_size |
| 109 | + mlp_params = gate_proj + up_proj + down_proj |
| 110 | + |
| 111 | + # Total per-layer |
| 112 | + layer_params = attn_params + mlp_params |
| 113 | + # Total params |
| 114 | + num_params = c.num_hidden_layers * layer_params + embedding_params |
| 115 | + return num_params * c.bytes_per_parameter / 1e9 |
| 116 | + |
| 117 | + @property |
| 118 | + def kv_cache_memory(self) -> float: |
| 119 | + """ |
| 120 | + Returns estimated KV cache memory in GB for GQA models. |
| 121 | +
|
| 122 | + Grouped Query Attention uses num_key_value_heads, which groups of Q heads share a K and V projection. |
| 123 | + num_key_value_heads < num_attention_heads, which reduces the KV Cache size. |
| 124 | + """ |
| 125 | + c = self.llm_config |
| 126 | + seq_len = self.seq_len or getattr(c, "max_seq_len", 2048) |
| 127 | + kv_cache_dtype_bytes = c.bytes_per_parameter |
| 128 | + kv_heads = c.num_key_value_heads |
| 129 | + |
| 130 | + total_bytes = ( |
| 131 | + self.batch_size |
| 132 | + * c.num_hidden_layers |
| 133 | + * 2 |
| 134 | + * kv_heads |
| 135 | + * seq_len |
| 136 | + * c.head_dim |
| 137 | + * kv_cache_dtype_bytes |
| 138 | + ) |
| 139 | + return total_bytes / 1e9 |
| 140 | + |
| 141 | + def _calc_attn_embed_params(self) -> tuple: |
| 142 | + """ |
| 143 | + Returns the embedding parameter count and attention parameter count for Llama-family (GQA) models. |
| 144 | + """ |
| 145 | + c = self.llm_config |
| 146 | + |
| 147 | + # Embedding parameters |
| 148 | + # assume tied embeddings unless tie_word_embeddings = False |
| 149 | + embedding_count = 1 if getattr(c, "tie_word_embeddings", True) else 2 |
| 150 | + embedding_params = embedding_count * c.vocab_size * c.hidden_size |
| 151 | + |
| 152 | + q_proj = c.hidden_size * c.hidden_size |
| 153 | + k_proj = c.hidden_size * (c.num_key_value_heads * c.head_dim) |
| 154 | + v_proj = c.hidden_size * (c.num_key_value_heads * c.head_dim) |
| 155 | + o_proj = c.hidden_size * c.hidden_size |
| 156 | + attn_params = q_proj + k_proj + v_proj + o_proj |
| 157 | + |
| 158 | + return embedding_params, attn_params |
| 159 | + |
| 160 | + |
| 161 | +class MixtureMemoryEstimator(LlamaMemoryEstimator): |
| 162 | + """ |
| 163 | + Estimator for Mixture-of-Experts (MoE) architectures (e.g., Mixtral, MoE Llama). |
| 164 | + Adds extra expert parallelism block parameter count to LlamaMemoryEstimator logic. |
| 165 | + """ |
| 166 | + |
| 167 | + @property |
| 168 | + def model_memory(self) -> float: |
| 169 | + """ |
| 170 | + Accounts for the increase in model parameters due to additional expert MLP blocks in MoE Models. |
| 171 | +
|
| 172 | + Returns the estimated memory size of the MoE Model (in GB). |
| 173 | + """ |
| 174 | + c = self.llm_config |
| 175 | + # Attention parameter count (Llama-style) |
| 176 | + embedding_params, attn_params = self._calc_attn_embed_params() |
| 177 | + |
| 178 | + # MoE MLP params per layer |
| 179 | + moe_params_per_layer = ( |
| 180 | + c.num_local_experts * 3 * c.hidden_size * c.intermediate_size |
| 181 | + ) |
| 182 | + total_params = ( |
| 183 | + c.num_hidden_layers * (attn_params + moe_params_per_layer) |
| 184 | + + embedding_params |
| 185 | + ) |
| 186 | + |
| 187 | + # Convert to GB |
| 188 | + return total_params * c.bytes_per_parameter / 1e9 |
| 189 | + |
| 190 | + |
| 191 | +def get_estimator(llm_config, **kwargs) -> MemoryEstimator: |
| 192 | + """ |
| 193 | + Extracts the correct estimator based on the defined parameters in the config.json |
| 194 | + See constants.py for LLMConfig parameters necessary for specific estimators. |
| 195 | + Uses MemoryEstimator as a fallback if parameters needed for GQA and MoE Architectures are missing. |
| 196 | +
|
| 197 | + Returns the appropriate MemoryEstimator based on the fields defined by the model's config.json (as represented by LLMConfig). |
| 198 | + """ |
| 199 | + if all( |
| 200 | + hasattr(llm_config, f) and getattr(llm_config, f) is not None |
| 201 | + for f in MOE_REQUIRED_FIELDS |
| 202 | + ): |
| 203 | + return MixtureMemoryEstimator(llm_config=llm_config, **kwargs) |
| 204 | + elif all( |
| 205 | + hasattr(llm_config, f) and getattr(llm_config, f) is not None |
| 206 | + for f in LLAMA_REQUIRED_FIELDS |
| 207 | + ): |
| 208 | + return LlamaMemoryEstimator(llm_config=llm_config, **kwargs) |
| 209 | + else: |
| 210 | + logger.warning( |
| 211 | + "Falling back to generic GPT estimator: required fields missing from config.json file in model." |
| 212 | + ) |
| 213 | + return MemoryEstimator(llm_config=llm_config, **kwargs) |
0 commit comments