File size: 20,399 Bytes
b647436 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 |
import json
from pathlib import Path
from typing import Union, Dict, Optional
def gqa_model_theoretical_flops(
config_path: Union[str, Path],
seq_len: int = 0,
gen_len: int = 1024,
batch_size: int = 1,
prefill_logits: str = "all", # "all" | "last" | "none"
) -> Dict[str, float]:
"""
Compute theoretical FLOPs for an LLM with GQA given its Hugging Face config.json.
Assumptions (dense Transformer, forward only):
- 2 FLOPs per multiply-add.
- Attention = dense GQA: Q & O project to d_model; K/V project to n_kv_heads * d_k
where d_k = d_model / n_heads.
- Attention core cost includes QK^T and softmax(QK^T) @ V.
- MLP = gated (SwiGLU-like): two "up" matmuls + one "down" matmul. (handles special
cases of llama-4 and gpt-oss)
- LM head (final logits) included; at prefill you can count logits for:
* "all": logits for every prompt token (matches HF's default forward outputs),
* "last": logits only for last prompt token (some gens do this),
* "none": if you never materialize logits at prefill.
At decode, logits are computed every step.
Returns (TFLOPs):
dict with detailed breakdown for prefill, decode, totals.
"""
# ---- load config ----
if "Ruyi" in config_path:
import re
pattern = re.compile(r'(\d+(?:\.\d+)?)\s*(?:B|billion)', re.IGNORECASE)
match = pattern.search(config_path)
config_path = config_path.replace(match.group(0), "7B")
cfg_path = Path(config_path)
if cfg_path.is_dir():
cfg_path = cfg_path / "config.json"
with open(cfg_path, "r") as f:
cfg = json.load(f)
if "gemma-3" in config_path:
import re
pattern = re.compile(r'(\d+(?:\.\d+)?)\s*(?:B|billion)', re.IGNORECASE)
match = pattern.search(config_path)
param_count = float(match.group(1))
if param_count >= 4:
cfg = cfg["text_config"]
cfg["vocab_size"] = 262208
if param_count == 4: cfg["num_attention_heads"] = 8; cfg["num_key_value_heads"] = 4
elif param_count == 12: cfg["num_attention_heads"] = 16; cfg["num_key_value_heads"] = 8
elif param_count == 27: cfg["num_attention_heads"] = 32; cfg["num_key_value_heads"] = 16
if "Llama-4" in config_path:
cfg = cfg["text_config"]
# ---- required hyperparams ----
d_model = int(cfg["hidden_size"])
n_layers = int(cfg.get("num_hidden_layers", cfg.get("n_layer"))) if "Ruyi" not in config_path else int(match.group(1)) * 4
n_heads = int(cfg.get("num_attention_heads", cfg.get("n_head")))
n_kv_heads = int(cfg.get("num_key_value_heads", n_heads))
if "Llama-4" in config_path: d_ff = cfg["intermediate_size_mlp"]
elif ("Qwen1.5" in config_path or "Qwen2-" in config_path) and "B-A" in config_path:
d_ff = cfg["intermediate_size"] + cfg["shared_expert_intermediate_size"]
else: d_ff = int(cfg.get("intermediate_size", cfg.get("ffn_hidden_size"))) # llama-4 uses intermediate_size_mlp for main mlp
vocab_size = int(cfg["vocab_size"])
# per-head dimension (assume divisible)
d_k = d_model // n_heads
kv_dim = n_kv_heads * d_k
B = batch_size
L = seq_len
T = gen_len
# ---- helpers (FLOPs, not TFLOPs) ----
# Projections per layer for a sequence of length L
# Q: 2 * B * L * d_model * d_model
# O: same
# K,V: 2 * B * L * d_model * kv_dim each
def proj_flops(L_tokens: int) -> int:
q = 2 * B * L_tokens * d_model * d_model
o = 2 * B * L_tokens * d_model * d_model
k = 2 * B * L_tokens * d_model * kv_dim
v = 2 * B * L_tokens * d_model * kv_dim
return q + k + v + o
# Attention core per layer
# Prefill (quadratic): QK^T + (softmax@V) ≈ 4 * B * n_heads * L^2 * d_k
# Decode (one step over cache length C): ≈ 4 * B * n_heads * C * d_k
def attn_core_prefill_flops(L_tokens: int) -> int:
return 4 * B * n_heads * (L_tokens ** 2) * d_k
def attn_core_decode_flops(cache_len: int) -> int:
return 4 * B * n_heads * cache_len * d_k
# MLP per layer
# Two up matmuls + one down: 2*B*L*d_model*d_ff + 2*B*L*d_model*d_ff + 2*B*L*d_ff*d_model = 6*B*L*d_model*d_ff
def mlp_flops(L_tokens: int) -> int:
# gpt-oss does not use gate function (6 → 4), registers per-expert intermediate size
if "gpt-oss" in config_path: return 4 * B * L_tokens * d_model * d_ff * int(cfg["num_experts_per_tok"])
# llama-4 use 2-layer mlp without gating on attn score, before the main mlp
elif "Llama-4" in config_path: return B * L_tokens * d_model * (6 * d_ff + 4 * int(cfg["intermediate_size"]))
else: return 6 * B * L_tokens * d_model * d_ff
# LM head (final linear to vocab) for N tokens: 2 * B * N * d_model * vocab_size
def lm_head_flops(num_tokens: int) -> int:
return 2 * B * num_tokens * d_model * vocab_size
# ---- prefill (length L) ----
proj_prefill_per_layer = proj_flops(L)
attn_prefill_per_layer = attn_core_prefill_flops(L)
mlp_prefill_per_layer = mlp_flops(L)
stack_prefill = n_layers * (proj_prefill_per_layer + attn_prefill_per_layer + mlp_prefill_per_layer)
if prefill_logits == "all":
lm_prefill = lm_head_flops(L)
elif prefill_logits == "last":
lm_prefill = lm_head_flops(1)
elif prefill_logits == "none":
lm_prefill = 0
else:
raise ValueError("prefill_logits must be one of {'all','last','none'}")
prefill_total = stack_prefill + lm_prefill
# ---- decode (T steps) ----
# For each step, projections/MLP are for 1 new token.
proj_decode_per_layer_per_step = proj_flops(1)
mlp_decode_per_layer_per_step = mlp_flops(1)
# Attention core sums over growing cache lengths: L, L+1, ..., L+T-1
# Sum_{t=0..T-1} 4 * B * n_heads * (L + t) * d_k = 4 * B * n_heads * d_k * (T*L + T*(T-1)/2)
attn_decode_per_layer_total = 4 * B * n_heads * d_k * (T * L + (T * (T - 1)) // 2)
stack_decode = n_layers * (
T * (proj_decode_per_layer_per_step + mlp_decode_per_layer_per_step) + attn_decode_per_layer_total
)
# Logits at each decode step
lm_decode = lm_head_flops(T)
decode_total = stack_decode + lm_decode
# ---- packing results (TFLOPs) ----
toT = lambda x: x / 1e12
results = {
# Inputs
"batch_size": B,
"seq_len": L,
"gen_len": T,
"hidden_size": d_model,
"num_layers": n_layers,
"num_heads": n_heads,
"num_kv_heads": n_kv_heads,
"intermediate_size": d_ff,
"vocab_size": vocab_size,
"prefill_logits_mode": prefill_logits,
# Prefill breakdown
"prefill_stack_TFLOPs": toT(stack_prefill),
"prefill_proj_TFLOPs": toT(n_layers * proj_prefill_per_layer),
"prefill_attn_core_TFLOPs": toT(n_layers * attn_prefill_per_layer),
"prefill_mlp_TFLOPs": toT(n_layers * mlp_prefill_per_layer),
"prefill_lm_head_TFLOPs": toT(lm_prefill),
"prefill_total_TFLOPs": toT(prefill_total),
# Decode breakdown
"decode_stack_TFLOPs": toT(stack_decode),
"decode_proj_TFLOPs": toT(n_layers * T * proj_decode_per_layer_per_step),
"decode_attn_core_TFLOPs": toT(n_layers * attn_decode_per_layer_total),
"decode_mlp_TFLOPs": toT(n_layers * T * mlp_decode_per_layer_per_step),
"decode_lm_head_TFLOPs": toT(lm_decode),
"decode_total_TFLOPs": toT(decode_total),
# Totals
"request_total_TFLOPs": toT(prefill_total + decode_total),
"avg_decode_TFLOPs_per_token": toT(decode_total / max(T, 1)),
}
return results
def mla_model_theoretical_flops(
config_path: Union[str, Path],
seq_len: int = 0,
gen_len: int = 1024,
batch_size: int = 1,
prefill_logits: str = "all", # "all" | "last" | "none"
attention_type: Optional[str] = None, # "mha" | "mla" | None (auto-detect)
mla_latents: Optional[int] = None,
mla_mode: str = "reuse", # "reuse" | "recompute"
) -> Dict[str, float]:
"""
Compute theoretical FLOPs (TFLOPs) for DeepSeek-R1 (or similar) inference.
Key points & assumptions (be sure to read):
- This function supports both classic dense Multi-Head Attention (MHA)
and DeepSeek's Multi-Head Latent Attention (MLA). MLA reduces the
attention core from O(L^2) to O(L * M) where M is the number of latent
tokens (per head or global depending on implementation). See DeepSeek-V2/V3 papers.
MLA also admits two execution schemes: 'reuse' (compute latent KV once at prefill
and reuse during decode) and 'recompute' (recompute / update latents per step).
The hardware analysis and community descriptions motivated these cost models.
- MoE MLP: we model a single shared expert (always executed) plus `num_experts_per_tok`
*activated* experts per token (as reported in the config). We expose separate
FLOP entries for shared vs activated experts.
- Projection FLOPs follow your previous convention: 2 FLOPs per multiply-add,
and we keep the same projection accounting for Q/K/V/O. The attention *core* cost
is replaced with MLA formulas when used.
- Because MLA variants differ in implementation details across repos, you can pass
`mla_latents` to set the latent length (if None a conservative default is used).
The default is chosen to reflect a moderate compression (an inferrable but tunable value).
- All counts are for forward-only inference, and result units are TFLOPs.
Parameters:
mla_latents: recommended to pass a locale-specific sensible value (e.g., 64, 128, 256).
If None, the function will pick a conservative default: min(256, max(1, seq_len // 16)).
mla_mode: "reuse" (default) counts the one-time cost to build latents at prefill and
then low-cost per-step decode attention against the smaller latent set.
"recompute" falls back to recomputing compressed latents per decode step
— yielding higher compute but lower memory footprint (useful to model
alternate execution strategies). See hardware-centric analysis.
"""
cfg_path = Path(config_path)
if cfg_path.is_dir():
cfg_path = cfg_path / "config.json"
with open(cfg_path, "r") as f:
cfg = json.load(f)
# ---- required hyperparams ----
d_model = int(cfg["hidden_size"])
n_layers = int(cfg["num_hidden_layers"])
n_heads = int(cfg["num_attention_heads"])
n_kv_heads = int(cfg.get("num_key_value_heads", n_heads))
d_ff = int(cfg.get("moe_intermediate_size", cfg.get("intermediate_size")))
vocab_size = int(cfg["vocab_size"])
# MoE-specific
n_experts_total = int(cfg.get("n_routed_experts", cfg.get("num_experts", cfg.get("num_local_experts", 0))))
n_shared_experts = int(cfg.get("n_shared_experts", cfg.get("n_shared_experts", 0)))
n_experts_per_tok = int(cfg.get("num_experts_per_tok", cfg.get("num_experts_per_tok", 0)))
# Detect/override attention type:
cfg_model_type = cfg.get("model_type", "").lower()
if attention_type is None:
# If model type contains deepseek or config contains MLA-related fields, default to mla
if "deepseek" in cfg_model_type or cfg.get("moa") or cfg.get("n_group") is not None:
attention_type = "mla"
else:
attention_type = "mha"
# MLA default latent length (tunable). MLA papers/reports show M << L; choose conservative default.
if mla_latents is None:
mla_latents = int(cfg.get("kv_lora_rank", max(1, min(256, max(1, seq_len // 16)))))
# per-head dimension (assume divisible)
d_k = d_model // n_heads
kv_dim = n_kv_heads * d_k
B = batch_size
L = seq_len
T = gen_len
# ---- helpers (FLOPs, NOT TFLOPs) ----
# Linear projections per layer for a sequence of length L_tokens.
# Keep original projection accounting for Q, O, K, V (this counts the input linear layers).
def proj_flops(L_tokens: int) -> int:
q = 2 * B * L_tokens * d_model * d_model # Wq : d_model x d_model
o = 2 * B * L_tokens * d_model * d_model # Wo : d_model x d_model (output projection)
# For K and V we keep the same "dense" projection accounting here. MLA adds separate
# compression costs which we model in attention_core_mla below.
k = 2 * B * L_tokens * d_model * kv_dim
v = 2 * B * L_tokens * d_model * kv_dim
return q + k + v + o
# Dense attention core (classic quadratic)
def attn_core_prefill_mha(L_tokens: int) -> int:
# approximate QK^T + softmax@V cost
return 4 * B * n_heads * (L_tokens ** 2) * d_k
def attn_core_decode_mha(cache_len: int) -> int:
return 4 * B * n_heads * cache_len * d_k
# MLA attention core (approximate): replace L^2 with L * M.
# We model two things:
# 1) core: Q @ K_latent^T and softmax@V_latent -> ~ 4 * B * n_heads * L * M * d_k
# 2) one-time compression cost at prefill to build the latent K/V (approximation).
# hardware analyses show there are two execution schemes: re-use (compress once) vs recompute.
# We approximate the one-time compression cost as: 2 * B * L * d_model * (mla_latents / max(1,L))
# which simplifies to ~ 2 * B * d_model * mla_latents (a compact, tunable approximation).
# See DeepSeek papers and hardware analysis for details.
def attn_core_prefill_mla(L_tokens: int) -> int:
M = mla_latents
core = 4 * B * n_heads * L_tokens * M * d_k
# one-time compress cost (approximation; tunable)
compress = int(2 * B * d_model * M)
return core + compress
def attn_core_decode_mla_reuse(L_tokens: int, T_steps: int) -> int:
# If latents are reused, each decode step attends Q (1 token) against latent keys size M:
# cost per step ~ 4 * B * n_heads * M * d_k
return 4 * B * n_heads * d_k * (T_steps * mla_latents)
def attn_core_decode_mla_recompute(L_tokens: int, T_steps: int) -> int:
# recomputing latents each step approximates back toward classic cost (worse-case).
# fall back to the MHA-like growing-cache sum as conservative upper bound:
return 4 * B * n_heads * d_k * (T_steps * L_tokens + (T_steps * (T_steps - 1)) // 2)
# MLP costs:
# Single expert (SwiGLU-like gated): approx 6 * B * L * d_model * d_ff
def single_expert_flops(L_tokens: int) -> int:
return 6 * B * L_tokens * d_model * d_ff
# MoE MLP breakdown: shared experts (n_shared_experts) executed every token
# plus activated experts (n_experts_per_tok) *per-token* (sparse routing).
# Note: some implementations add extra routing overhead; we ignore the small routing bookkeeping cost here.
def moe_mlp_flops_shared(L_tokens: int) -> int:
# FLOPs for shared (always executed). If config says n_shared_experts>1, multiply accordingly.
return n_shared_experts * single_expert_flops(L_tokens)
def moe_mlp_flops_activated(L_tokens: int) -> int:
# Activated experts per token: each token runs n_experts_per_tok experts (sparse).
return n_experts_per_tok * single_expert_flops(L_tokens)
# LM head
def lm_head_flops(num_tokens: int) -> int:
return 2 * B * num_tokens * d_model * vocab_size
# ---- PREFILL (length L) ----
proj_prefill_per_layer = proj_flops(L)
if attention_type == "mha":
attn_prefill_per_layer = attn_core_prefill_mha(L)
# no extra MLA compress cost
mla_extra_prefill_per_layer = 0
elif attention_type == "mla":
attn_prefill_per_layer = attn_core_prefill_mla(L)
# the compression cost is included in attn_core_prefill_mla as 'compress' term
mla_extra_prefill_per_layer = max(0, attn_prefill_per_layer - (4 * B * n_heads * (L ** 2) * d_k))
else:
raise ValueError("attention_type must be one of {'mha','mla'}")
# MLP (MoE)
mlp_prefill_shared_per_layer = moe_mlp_flops_shared(L)
mlp_prefill_activated_per_layer = moe_mlp_flops_activated(L)
mlp_prefill_per_layer = mlp_prefill_shared_per_layer + mlp_prefill_activated_per_layer
stack_prefill = n_layers * (proj_prefill_per_layer + attn_prefill_per_layer + mlp_prefill_per_layer)
if prefill_logits == "all":
lm_prefill = lm_head_flops(L)
elif prefill_logits == "last":
lm_prefill = lm_head_flops(1)
elif prefill_logits == "none":
lm_prefill = 0
else:
raise ValueError("prefill_logits must be one of {'all','last','none'}")
prefill_total = stack_prefill + lm_prefill
# ---- DECODE (T steps) ----
proj_decode_per_layer_per_step = proj_flops(1)
mlp_decode_per_layer_per_step_shared = moe_mlp_flops_shared(1)
mlp_decode_per_layer_per_step_activated = moe_mlp_flops_activated(1)
mlp_decode_per_layer_per_step = mlp_decode_per_layer_per_step_shared + mlp_decode_per_layer_per_step_activated
if attention_type == "mha":
# attention grows with cache: L, L+1, ..., L+T-1
attn_decode_per_layer_total = 4 * B * n_heads * d_k * (T * L + (T * (T - 1)) // 2)
mla_extra_decode_term = 0
else: # mla
if mla_mode == "reuse":
attn_decode_per_layer_total = attn_core_decode_mla_reuse(L, T)
mla_extra_decode_term = 0 # compression cost already accounted in prefill
elif mla_mode == "recompute":
attn_decode_per_layer_total = attn_core_decode_mla_recompute(L, T)
# recompute implies we pay full compress-like cost in decode as well;
# approximate by adding the same compress cost per layer per decode (conservative)
per_step_compress = int(2 * B * d_model * mla_latents)
mla_extra_decode_term = n_layers * (per_step_compress * T)
else:
raise ValueError("mla_mode must be one of {'reuse','recompute'}")
stack_decode = n_layers * (
T * (proj_decode_per_layer_per_step + mlp_decode_per_layer_per_step) + attn_decode_per_layer_total
) + mla_extra_decode_term
lm_decode = lm_head_flops(T)
decode_total = stack_decode + lm_decode
# ---- pack results (TFLOPs) ----
toT = lambda x: x / 1e12
results = {
# Inputs / config readout
"batch_size": B,
"seq_len": L,
"gen_len": T,
"hidden_size": d_model,
"num_layers": n_layers,
"num_heads": n_heads,
"num_kv_heads": n_kv_heads,
"intermediate_size": d_ff,
"vocab_size": vocab_size,
"num_experts_total": n_experts_total,
"num_shared_experts": n_shared_experts,
"num_experts_per_tok": n_experts_per_tok,
"attention_type": attention_type,
"mla_latents": mla_latents if attention_type == "mla" else None,
"mla_mode": mla_mode if attention_type == "mla" else None,
"prefill_logits_mode": prefill_logits,
# Prefill breakdown
"prefill_stack_TFLOPs": toT(stack_prefill),
"prefill_proj_TFLOPs": toT(n_layers * proj_prefill_per_layer),
"prefill_attn_core_TFLOPs": toT(n_layers * attn_prefill_per_layer),
"prefill_mlp_shared_TFLOPs": toT(n_layers * mlp_prefill_shared_per_layer),
"prefill_mlp_activated_TFLOPs": toT(n_layers * mlp_prefill_activated_per_layer),
"prefill_mlp_TFLOPs": toT(n_layers * mlp_prefill_per_layer),
"prefill_lm_head_TFLOPs": toT(lm_prefill),
"prefill_total_TFLOPs": toT(prefill_total),
# Decode breakdown
"decode_stack_TFLOPs": toT(stack_decode),
"decode_proj_TFLOPs": toT(n_layers * T * proj_decode_per_layer_per_step),
"decode_attn_core_TFLOPs": toT(n_layers * attn_decode_per_layer_total),
"decode_mlp_shared_TFLOPs": toT(n_layers * T * mlp_decode_per_layer_per_step_shared),
"decode_mlp_activated_TFLOPs": toT(n_layers * T * mlp_decode_per_layer_per_step_activated),
"decode_mlp_TFLOPs": toT(n_layers * T * mlp_decode_per_layer_per_step),
"decode_lm_head_TFLOPs": toT(lm_decode),
"decode_total_TFLOPs": toT(decode_total),
# Totals
"request_total_TFLOPs": toT(prefill_total + decode_total),
"avg_decode_TFLOPs_per_token": toT(decode_total / max(T, 1)),
}
return results |