Files
Xiaoze Fan 56e7f3ae38 feat(quant)!: compress bf16 weights to df11 at load with --online-quant (#891)
* feat(quant)!: compress bf16 weights to df11 at load with --online-quant

GLM-4.7 no longer hardwires DF11 attention and embedding; it serves them
bf16 unless --online-quant df11 is passed, and FTW files converted with the
old DF11 tensors must be reconverted.

* fix(quant): skip layers --online-quant cannot convert instead of failing
2026-11-11 13:24:46 -08:00

195 lines
8.5 KiB
Python

from __future__ import annotations
from typing import Dict
import torch
from freetoken.core import get_global_ctx
from freetoken.distributed import DistributedCommunicator, get_tp_info
from freetoken.utils import div_ceil, init_logger, nvtx_annotate
from .base import BaseOP
from .quantization import LayerKind, QuantConfig, quant_method_for
logger = init_logger(__name__)
class VocabParallelEmbedding(BaseOP):
"""The vocab table, sharded over TP: its ``quant_method`` owns the table, looks rows up, and serves a tied LM head's product."""
quant_layer_kind = LayerKind.EMBEDDING
def __init__(
self,
num_embeddings: int,
embedding_dim: int,
embed_scale: float | None = None,
*,
quant_config: QuantConfig | None = None,
prefix: str = "",
):
super().__init__()
tp_info = get_tp_info()
tp_rank = tp_info.rank
self.tp_size = tp_info.size
self.num_embeddings = num_embeddings
self.num_embeddings_tp = div_ceil(num_embeddings, self.tp_size)
start_idx = self.num_embeddings_tp * tp_rank
finish_idx = min(start_idx + self.num_embeddings_tp, num_embeddings)
self.vocab_range = (start_idx, finish_idx - start_idx)
self.embedding_dim = embedding_dim
self.prefix = prefix
# the LM head subclass is a linear and builds its own method
self.quant_method = None
if self.quant_layer_kind is LayerKind.EMBEDDING:
self.quant_method = quant_method_for(quant_config, self, prefix)
self.quant_method.create_weights(self)
# Gemma scales embeddings by sqrt(hidden_size). The scale is materialized in
# the weight dtype (bf16) to match HF, which downcasts the scalar. The GPU
# scalar is built lazily (model __init__ runs on the meta device) and cached
# so it is not reallocated inside a captured CUDA graph.
self._embed_scale = embed_scale
self._embed_scale_t: torch.Tensor | None = None
self._comm = DistributedCommunicator()
@nvtx_annotate("Embedding")
def forward(self, x: torch.Tensor) -> torch.Tensor:
y = self.quant_method.embed(self, x)
if self.tp_size > 1:
y = self._comm.all_reduce(y)
if self._embed_scale is not None:
if self._embed_scale_t is None:
self._embed_scale_t = torch.tensor(
self._embed_scale, dtype=y.dtype, device=y.device
)
y = y * self._embed_scale_t
return y
class ParallelLMHead(VocabParallelEmbedding):
"""The head is a linear layer over the vocab shard: its weights come from ``quant_method``
unless they are tied to the input embedding."""
quant_layer_kind = LayerKind.LINEAR
def __init__(
self,
num_embeddings: int,
embedding_dim: int,
bias: bool = False,
tie_word_embeddings: bool = False,
tied_embedding: VocabParallelEmbedding | None = None,
*,
quant_config: QuantConfig | None = None,
prefix: str = "",
):
super().__init__(num_embeddings, embedding_dim)
self.has_bias = bias
self.prefix = prefix
self.tied_embedding = tied_embedding
assert (tied_embedding is not None) == tie_word_embeddings
if tied_embedding is not None and tied_embedding.quant_method.online:
# the head multiplies by the whole table every step, which an online table would decode in full each time
logger.info("--online-quant leaves %s as stored: the LM head is tied to it", tied_embedding.prefix)
tied_embedding.quant_method = quant_method_for(None, tied_embedding, tied_embedding.prefix)
tied_embedding.quant_method.create_weights(tied_embedding)
self.in_features = embedding_dim
self.out_features = self.num_embeddings_tp
self.output_sizes = (self.num_embeddings_tp,)
if tied_embedding is None:
self.quant_method = quant_method_for(quant_config, self, prefix)
self.quant_method.create_weights(self)
self.bias = torch.empty(self.num_embeddings_tp) if bias else None
# set by keep_rows: a head over a subset of the rows, and the token id of each of its rows
self.kept_rows: tuple[ParallelLMHead, torch.Tensor] | None = None
def finalize(self) -> None:
if self.quant_method is not None:
self.quant_method.finalize(self)
def keep_rows(self, ids: torch.Tensor, quant_config: QuantConfig | None) -> None:
"""Set ``kept_rows`` to a head with its own weights over only the rows ``ids``, repeated up to a multiple of 128 rows.
Call after the weights load and before ``finalize_quant``: a kernel's finalize can rewrite the weights so that rows cannot be picked, and ``finalize_quant`` also finalizes the new head.
"""
# scaled_mm needs N % 16, b12x N % 8, and the sm_120 GEMV's fast path N % 128; a repeated row maps to the same token
ids = ids[torch.arange(div_ceil(ids.numel(), 128) * 128) % ids.numel()]
rows = ids.numel()
with torch.device("meta"):
head = ParallelLMHead(rows, self.embedding_dim, self.has_bias, quant_config=quant_config, prefix=self.prefix)
full = self.state_dict()
ids = ids.to(next(iter(full.values())).device)
for name, want in head.state_dict().items():
have = full[name]
if want.dim() == 0:
kept = have.clone()
elif want.shape[0] >= rows and have.shape[1:] == want.shape[1:] and have.shape[0] - self.num_embeddings == want.shape[0] - rows:
kept = have.index_select(0, ids)
if want.shape[0] > rows:
# ggml's kernels read past the last row and need zeros there
kept = torch.cat([kept, kept.new_zeros(want.shape[0] - rows, *kept.shape[1:])])
else:
raise NotImplementedError(f"{self.prefix}.{name} is not stored by row (shape {list(have.shape)}), so the head cannot keep a subset of its rows")
setattr(head, name, kept)
self.kept_rows = head, ids
def load_state_dict(
self,
state_dict: Dict[str, torch.Tensor],
*,
prefix: str = "",
_internal: bool = False,
) -> None:
if not self.tied_embedding:
return super().load_state_dict(state_dict, prefix=prefix, _internal=_internal)
else:
# pop the lm_head.weights and lm_head.bias if they exist
possible_weight = f"{prefix}.weight"
possible_bias = f"{prefix}.bias"
if possible_weight in state_dict:
state_dict.pop(possible_weight)
if possible_bias in state_dict:
state_dict.pop(possible_bias)
def state_dict(
self,
*,
prefix: str = "",
result: Dict[str, torch.Tensor] | None = None,
) -> Dict[str, torch.Tensor]:
if not self.tied_embedding:
return super().state_dict(prefix=prefix, result=result)
return {} if result is None else result
@nvtx_annotate("LMHead")
def forward(self, x: torch.Tensor, *, export_hidden: bool = True) -> torch.Tensor:
"""``export_hidden=False``: the model hands its own final hidden to ``batch.features``."""
ctx = get_global_ctx()
batch = ctx.batch
bs = batch.size
if export_hidden and (features := batch.features) is not None:
features.final_hidden(x)
if batch.is_prefill:
indices = batch.attn_metadata.get_last_indices(bs)
x = x[indices].contiguous()
del indices
return self.logits(x)
def logits(self, x: torch.Tensor) -> torch.Tensor:
"""``num_embeddings`` logits for each row of ``x``; reads no batch state."""
if self.tied_embedding is not None:
logits = self.tied_embedding.quant_method.logits(self.tied_embedding, x, self.bias)
else:
logits = self.quant_method.apply(self, x)
if self.tp_size == 1:
return logits
input_shape = logits.shape
output_tensor = self._comm.all_gather(logits)
# keyed on rows, not batch size: a verify batch has several rows per request
if input_shape[0] == 1:
return output_tensor.view(1, -1)[:, : self.num_embeddings]
output_tensor = output_tensor.view((self.tp_size,) + input_shape)
output_tensor = output_tensor.permute(1, 0, 2).contiguous()
output_tensor = output_tensor.reshape(input_shape[:1] + (self.tp_size * input_shape[1],))
return output_tensor[:, : self.num_embeddings]