supermix-v80 / src /mimomix_decoding.py
Kai9987kai's picture
Supermix v80: 15.3M-parameter physics and arithmetic solver, solver-verified corpus
fd32dda verified
Raw
History Blame Contribute Delete
16.3 kB
"""MTP self-speculative decoding for MiMoMix.
The Multi-Token Prediction depths trained in :mod:`mimomix_core` are reused at
inference as a *draft model that costs one block each*, which is the trick MiMo
uses to roughly triple output throughput without a second model in memory.
The loop implemented here is the standard draft/verify schedule specialised to
greedy decoding, where acceptance has an exact form:
accept a drafted token iff it equals the trunk's own argmax at that position
On the first mismatch the trunk's argmax is emitted instead and the rest of the
draft is discarded. This makes the emitted sequence **bit-identical** to plain
autoregressive greedy decoding -- speculation buys throughput and changes
nothing else. :func:`assert_greedy_equivalence` checks exactly that, and the
test-suite runs it on random models.
Two implementation details are load-bearing and easy to get wrong:
* **Cache rollback.** Rejecting ``r`` tokens means the KV entries written for
them must go. Under sliding-window attention a cache trimmed to exactly
``window`` has already dropped keys that rollback brings back into range, so
the decoder asks the model for ``cache_slack = draft_length`` extra entries.
* **Block-independence.** Verification feeds several positions at once. That is
only equivalent to one-at-a-time decoding if the model's per-position output
does not depend on what else is in the block. The adaptive thinking core
makes a *batch-level* halting decision, so speculative decoding refuses to
run with ``adaptive_thinking=True`` rather than quietly breaking the
guarantee.
"""
from __future__ import annotations
import time
from dataclasses import dataclass, asdict, field
from typing import Dict, List, Optional, Sequence, Tuple
import torch
from mimomix_core import MiMoMixModel
__all__ = [
"DecodeStats",
"GenerationResult",
"greedy_generate",
"speculative_generate",
"assert_greedy_equivalence",
"hybrid_cache_footprint",
"trim_past",
]
PastKV = List[Optional[Tuple[torch.Tensor, torch.Tensor]]]
@dataclass
class DecodeStats:
"""Throughput accounting for one generation call."""
mode: str
new_tokens: int = 0
#: forward passes through the full trunk (excluding the prefill)
verify_forwards: int = 0
prefill_forwards: int = 1
drafted_tokens: int = 0
accepted_draft_tokens: int = 0
seconds: float = 0.0
@property
def acceptance_length(self) -> float:
"""Mean tokens committed per trunk forward -- the headline MTP number.
Plain greedy decoding scores exactly ``1.0``. MiMo reports up to 3.6
with three MTP layers on a real checkpoint; an untrained toy model will
score near 1 because its drafts are noise, and that is the correct
behaviour, not a bug.
The prompt prefill both consumes a trunk forward and produces the first
generated token. Decode throughput deliberately excludes that common
setup cost from *both* sides of the ratio: ``verify_forwards`` does not
include prefill, so the token produced by prefill must not be counted in
the numerator either. Mixing those conventions makes greedy score
``N / (N - 1)`` and can put speculation above its theoretical
``draft_length + 1`` maximum.
"""
if self.verify_forwards == 0:
return 1.0 if self.mode == "greedy" and self.new_tokens > 0 else 0.0
return self.decoding_tokens / self.verify_forwards
@property
def decoding_tokens(self) -> int:
"""Generated tokens attributable to post-prefill decode forwards."""
prefill_token = 1 if self.prefill_forwards > 0 and self.new_tokens > 0 else 0
return max(0, int(self.new_tokens) - prefill_token)
@property
def acceptance_rate(self) -> float:
"""Fraction of speculative tokens that survived verification."""
if self.drafted_tokens == 0:
return 0.0
return self.accepted_draft_tokens / self.drafted_tokens
@property
def tokens_per_second(self) -> float:
if self.seconds <= 0.0:
return 0.0
return self.new_tokens / self.seconds
def to_dict(self) -> Dict[str, object]:
payload = asdict(self)
payload["decoding_tokens"] = self.decoding_tokens
payload["acceptance_length"] = round(self.acceptance_length, 4)
payload["acceptance_rate"] = round(self.acceptance_rate, 4)
payload["tokens_per_second"] = round(self.tokens_per_second, 3)
return payload
@dataclass
class GenerationResult:
sequences: torch.Tensor
new_tokens: torch.Tensor
stats: DecodeStats
telemetry: Dict[str, object] = field(default_factory=dict)
def trim_past(past: Optional[Sequence], drop: int) -> Optional[PastKV]:
"""Remove the last ``drop`` cached positions from every layer."""
if past is None or drop <= 0:
return None if past is None else list(past)
trimmed: PastKV = []
for entry in past:
if entry is None or entry[0].numel() == 0:
trimmed.append(entry)
continue
keys, values = entry
length = keys.shape[2]
keep = max(0, length - drop)
trimmed.append((keys[:, :, :keep], values[:, :, :keep]))
return trimmed
def _reject_adaptive(adaptive_thinking: bool) -> None:
if adaptive_thinking:
raise ValueError(
"speculative decoding requires a block-independent target model; "
"the adaptive thinking core halts on a batch-level statistic, so "
"pass adaptive_thinking=False (a fixed cycle budget) instead"
)
@torch.no_grad()
def greedy_generate(
model: MiMoMixModel,
input_ids: torch.Tensor,
max_new_tokens: int = 16,
eos_token_id: Optional[int] = None,
thinking_cycles: Optional[int] = None,
adaptive_thinking: bool = False,
) -> GenerationResult:
"""Reference one-token-at-a-time greedy decoding.
This is the correctness oracle for :func:`speculative_generate`.
"""
model.eval()
device = input_ids.device
started = time.perf_counter()
out = model(
input_ids,
use_cache=True,
thinking_cycles=thinking_cycles,
adaptive_thinking=adaptive_thinking,
return_mtp=False,
past_length=0,
)
past = out.past_key_values
position = int(input_ids.shape[1])
token = out.logits[:, -1].argmax(dim=-1, keepdim=True)
emitted: List[torch.Tensor] = []
stats = DecodeStats(mode="greedy")
finished = torch.zeros(input_ids.shape[0], dtype=torch.bool, device=device)
for _ in range(max_new_tokens):
if eos_token_id is not None and bool(finished.any()):
# Keep already-finished batch rows pinned to EOS while unfinished
# rows continue. For batch size one the loop exits immediately;
# for larger batches this gives every row standard stop semantics
# without returning ragged tensors.
token = torch.where(
finished.unsqueeze(1),
torch.full_like(token, int(eos_token_id)),
token,
)
emitted.append(token)
stats.new_tokens += 1
if eos_token_id is not None:
finished = finished | token.squeeze(1).eq(int(eos_token_id))
if bool(finished.all()):
break
if stats.new_tokens >= max_new_tokens:
break
step = model(
token,
past_key_values=past,
use_cache=True,
thinking_cycles=thinking_cycles,
adaptive_thinking=adaptive_thinking,
return_mtp=False,
past_length=position,
)
stats.verify_forwards += 1
past = step.past_key_values
position += 1
token = step.logits[:, -1].argmax(dim=-1, keepdim=True)
stats.seconds = time.perf_counter() - started
new_tokens = torch.cat(emitted, dim=1) if emitted else input_ids.new_zeros((input_ids.shape[0], 0))
return GenerationResult(
sequences=torch.cat([input_ids, new_tokens], dim=1),
new_tokens=new_tokens,
stats=stats,
telemetry=out.telemetry,
)
@torch.no_grad()
def speculative_generate(
model: MiMoMixModel,
input_ids: torch.Tensor,
max_new_tokens: int = 16,
eos_token_id: Optional[int] = None,
thinking_cycles: Optional[int] = None,
adaptive_thinking: bool = False,
draft_length: Optional[int] = None,
) -> GenerationResult:
"""Greedy decoding accelerated by the model's own MTP depths.
Emits exactly what :func:`greedy_generate` emits, using fewer trunk
forwards whenever the draft is right.
"""
_reject_adaptive(adaptive_thinking)
model.eval()
device = input_ids.device
batch = int(input_ids.shape[0])
max_draft = len(model.mtp_modules) if draft_length is None else int(draft_length)
max_draft = max(0, min(max_draft, len(model.mtp_modules)))
started = time.perf_counter()
prefill = model(
input_ids,
use_cache=True,
thinking_cycles=thinking_cycles,
adaptive_thinking=False,
return_mtp=False,
cache_slack=max_draft,
past_length=0,
)
past = prefill.past_key_values
committed_length = int(input_ids.shape[1])
# The trunk's own argmax at the last prompt position: exact, not a draft.
token = prefill.logits[:, -1].argmax(dim=-1, keepdim=True)
trunk_state = prefill.trunk_hidden[:, -1:]
emitted: List[torch.Tensor] = [token]
stats = DecodeStats(mode="speculative")
stats.new_tokens = 1
finished = torch.zeros(batch, dtype=torch.bool, device=device)
if eos_token_id is not None:
finished = token.squeeze(1).eq(int(eos_token_id))
while stats.new_tokens < max_new_tokens and not bool(finished.all()):
draft = model.propose_draft(trunk_state, token, position=committed_length - 1)
# Do not verify draft positions that cannot fit in the caller's output
# budget. One slot is reserved for the target model's bonus/correction
# token, so all accounting describes tokens that can actually be
# returned.
remaining = max_new_tokens - stats.new_tokens
draft_budget = min(max_draft, max(0, remaining - 1))
if draft_budget < draft.shape[1]:
draft = draft[:, :draft_budget]
block = torch.cat([token, draft], dim=1) if draft.numel() else token
block_len = int(block.shape[1])
n_draft = block_len - 1
step = model(
block,
past_key_values=past,
use_cache=True,
thinking_cycles=thinking_cycles,
adaptive_thinking=False,
return_mtp=False,
cache_slack=max_draft,
past_length=committed_length,
)
stats.verify_forwards += 1
stats.drafted_tokens += n_draft * batch
target = step.logits.argmax(dim=-1) # (B, block_len)
# Accept the longest prefix that every *unfinished* batch row agrees
# with. Batching forces a common accept length; per-row divergence just
# costs speed. Rows that reached EOS inside the block are ignored at
# later positions and are pinned to EOS in the returned tensor.
accepted = 0
verification_finished = finished.clone()
for index in range(n_draft):
candidate = block[:, index + 1]
matches = candidate.eq(target[:, index])
if not bool(matches[~verification_finished].all()):
break
accepted += 1
if eos_token_id is not None:
verification_finished = verification_finished | candidate.eq(int(eos_token_id))
if bool(verification_finished.all()):
break
stats.accepted_draft_tokens += accepted * batch
committed: List[torch.Tensor] = []
for index in range(accepted):
candidate = block[:, index + 1 : index + 2]
if eos_token_id is not None:
candidate = torch.where(
finished.unsqueeze(1),
torch.full_like(candidate, int(eos_token_id)),
candidate,
)
finished = finished | candidate.squeeze(1).eq(int(eos_token_id))
committed.append(candidate)
# If every row ended on an accepted draft token, greedy decoding would
# stop there. Do not append the block's bonus token after EOS.
bonus: Optional[torch.Tensor] = None
if not bool(finished.all()):
bonus = target[:, accepted : accepted + 1]
if eos_token_id is not None:
bonus = torch.where(
finished.unsqueeze(1),
torch.full_like(bonus, int(eos_token_id)),
bonus,
)
finished = finished | bonus.squeeze(1).eq(int(eos_token_id))
committed.append(bonus)
emitted.extend(committed)
committed_length += accepted + 1
rejected = n_draft - accepted
if rejected > 0:
past = trim_past(step.past_key_values, rejected)
else:
past = step.past_key_values
stats.new_tokens += len(committed)
if bonus is None:
break
trunk_state = step.trunk_hidden[:, accepted : accepted + 1]
token = bonus
stats.seconds = time.perf_counter() - started
new_tokens = torch.cat(emitted, dim=1)[:, :max_new_tokens]
stats.new_tokens = int(new_tokens.shape[1])
return GenerationResult(
sequences=torch.cat([input_ids, new_tokens], dim=1),
new_tokens=new_tokens,
stats=stats,
telemetry=prefill.telemetry,
)
def assert_greedy_equivalence(
model: MiMoMixModel,
input_ids: torch.Tensor,
max_new_tokens: int = 16,
thinking_cycles: Optional[int] = None,
) -> Dict[str, object]:
"""Run both decoders and require identical output. Raises on divergence."""
reference = greedy_generate(
model, input_ids, max_new_tokens=max_new_tokens, thinking_cycles=thinking_cycles
)
fast = speculative_generate(
model, input_ids, max_new_tokens=max_new_tokens, thinking_cycles=thinking_cycles
)
if reference.new_tokens.shape != fast.new_tokens.shape or not torch.equal(
reference.new_tokens, fast.new_tokens
):
raise AssertionError(
"speculative decoding diverged from greedy decoding\n"
f" greedy: {reference.new_tokens.tolist()}\n"
f" speculative: {fast.new_tokens.tolist()}"
)
return {
"tokens": int(reference.new_tokens.shape[1]),
"greedy_forwards": reference.stats.verify_forwards,
"speculative_forwards": fast.stats.verify_forwards,
"acceptance_length": round(fast.stats.acceptance_length, 4),
"acceptance_rate": round(fast.stats.acceptance_rate, 4),
"forward_reduction": round(
1.0 - (fast.stats.verify_forwards / max(1, reference.stats.verify_forwards)), 4
),
}
def hybrid_cache_footprint(model: MiMoMixModel, sequence_length: int) -> Dict[str, object]:
"""KV-cache entries a hybrid layout holds versus an all-global one.
This is the arithmetic behind the "hybrid attention shrinks the KV cache"
claim, evaluated for *this* model's layout. It counts cache entries, not
bytes, and assumes the cache is already at steady state.
"""
window = int(model.config.sliding_window)
per_layer: List[int] = []
for kind in model.layout:
per_layer.append(sequence_length if kind == "global" else min(window, sequence_length))
hybrid_total = sum(per_layer)
dense_total = sequence_length * len(model.layout)
return {
"sequence_length": int(sequence_length),
"sliding_window": window,
"layout": list(model.layout),
"per_layer_entries": per_layer,
"hybrid_entries": int(hybrid_total),
"all_global_entries": int(dense_total),
"reduction_factor": round(dense_total / hybrid_total, 4) if hybrid_total else 0.0,
"saved_fraction": round(1.0 - (hybrid_total / dense_total), 4) if dense_total else 0.0,
}