A symbol scan over every tracked file found 123 top-level definitions in vllm/ with zero references, including references inside strings. All of them came from features this fork deleted earlier. - Delete dead attention helpers: MLADCPManager, split_prefill_chunks, the reshape helpers, compute_varlen_chunk_metadata, the LINEAR backend's unused BailingLinearAttentionBackend, and the sparse flash-attn entry points. - Delete the CPU and AMX op wrappers from _custom_ops.py: cpu_mla_decode, cpu_mla_extend, cpu_mla_get_scheduler_metadata, gather_mla_context_cache_cpu. - Delete dead quantization helpers: awq_pack, gptq_pack, gptq_quantize_weights, get_and_maybe_dequant_weights, marlin_quant_fp8_torch, the rand_marlin_weight_*_like pair, ocp_mx_weight_dtype_and_rows, get_fp8_block_weight_scale, input_to_float8. - Delete dead model helpers: resolve_visual_encoder_outputs, run_dp_sharded_vision_model, fast_topk, get_layer_index, is_pp_missing_parameter, skip_spec_layers, KimiK3MergedQKVGateLinear, selective_scan_fn, causal_conv1d_update_torch. - Delete dead utils: format_kib, format_mib, chunk_list, prev_power_of_2, split_host_port, join_host_port, make_zmq_path, make_async_with_semaphore, is_meta_module, is_non_overlapping_and_dense, supports_xpu_graph, create_kv_caches_with_random*, and the unused has_* capability probes. - Delete the unreachable deep-gemm mega-MoE warmup path, which imported vllm.models.deepseek_v4.nvidia.model and vllm.models.kimi_k3.nvidia.model, neither of which exists anymore. - Drop the imports and the two now-empty guards left behind. Verified statically: every changed file compiles, no unused import remains, and no deleted name is referenced in vllm, tests or csrc. The runtime gates still need the Python environment rebuilt. Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
129 lines
3.7 KiB
Python
129 lines
3.7 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Accessors for the optional ``b12x`` package."""
|
|
|
|
import importlib
|
|
import importlib.util
|
|
from collections.abc import Callable, Hashable, Iterable
|
|
from dataclasses import dataclass, fields, is_dataclass
|
|
from types import ModuleType
|
|
from typing import Any
|
|
|
|
import torch
|
|
|
|
# Logical elements and encoded bytes per block.
|
|
B12X_BLOCK_CODECS = {"iq2_xs": (256, 74), "iq2_xxs": (256, 66), "q8_0": (32, 34)}
|
|
|
|
|
|
@dataclass(frozen=True)
|
|
class B12xWarmupUnit:
|
|
name: str
|
|
key: Hashable
|
|
compile: Callable[[], None]
|
|
|
|
|
|
_HAS_B12X = importlib.util.find_spec("b12x") is not None
|
|
|
|
|
|
def _import_submodule(module_name: str) -> ModuleType | None:
|
|
if not _HAS_B12X:
|
|
return None
|
|
try:
|
|
return importlib.import_module(module_name)
|
|
except (ImportError, ModuleNotFoundError):
|
|
return None
|
|
|
|
|
|
_B12X_SUBMODULES = {
|
|
module_name: _import_submodule(module_name)
|
|
for module_name in (
|
|
"b12x.attention.paged",
|
|
"b12x.gemm.blockscaled",
|
|
# TODO: Remove once B12X exposes the scale-swizzle API publicly.
|
|
"b12x._lib.intrinsics",
|
|
"b12x.gemm.mxfp8_linear",
|
|
"b12x.gemm.tensor_fp8_linear",
|
|
"b12x.moe.fused_moe",
|
|
)
|
|
}
|
|
|
|
|
|
def _get_submodule(module_name: str) -> ModuleType | None:
|
|
return _B12X_SUBMODULES.get(module_name)
|
|
|
|
|
|
def get_b12x_blockscaled() -> ModuleType | None:
|
|
return _get_submodule("b12x.gemm.blockscaled")
|
|
|
|
|
|
def get_b12x_intrinsics() -> ModuleType | None:
|
|
return _get_submodule("b12x._lib.intrinsics")
|
|
|
|
|
|
def get_b12x_mxfp8_linear() -> ModuleType | None:
|
|
return _get_submodule("b12x.gemm.mxfp8_linear")
|
|
|
|
|
|
def get_b12x_tensor_fp8_linear() -> ModuleType | None:
|
|
return _get_submodule("b12x.gemm.tensor_fp8_linear")
|
|
|
|
|
|
def get_b12x_fused_moe() -> ModuleType | None:
|
|
return _get_submodule("b12x.moe.fused_moe")
|
|
|
|
|
|
def b12x_warmup_token_counts(
|
|
*,
|
|
max_tokens: int,
|
|
cudagraph_capture_sizes: Iterable[int] = (),
|
|
) -> tuple[int, ...]:
|
|
# B12X deduplicates shapes that select the same internal kernel policy.
|
|
# Keep the complete serving shape set here rather than duplicating its
|
|
# policy-selection heuristics in vLLM.
|
|
counts = {1}
|
|
counts.update(int(size) for size in cudagraph_capture_sizes if int(size) > 0)
|
|
if int(max_tokens) > 0:
|
|
counts.add(int(max_tokens))
|
|
return tuple(sorted(counts))
|
|
|
|
|
|
def _same_packed_layout(current: Any, replacement: Any) -> bool:
|
|
if type(current) is not type(replacement):
|
|
return False
|
|
if isinstance(current, torch.Tensor):
|
|
return (
|
|
current.shape == replacement.shape
|
|
and current.stride() == replacement.stride()
|
|
and current.dtype == replacement.dtype
|
|
and current.device == replacement.device
|
|
)
|
|
if is_dataclass(current):
|
|
return all(
|
|
_same_packed_layout(
|
|
getattr(current, field.name),
|
|
getattr(replacement, field.name),
|
|
)
|
|
for field in fields(current)
|
|
)
|
|
return bool(current == replacement)
|
|
|
|
|
|
def _copy_packed_tensors(current: Any, replacement: Any) -> None:
|
|
if isinstance(current, torch.Tensor):
|
|
current.copy_(replacement)
|
|
elif is_dataclass(current):
|
|
for field in fields(current):
|
|
_copy_packed_tensors(
|
|
getattr(current, field.name),
|
|
getattr(replacement, field.name),
|
|
)
|
|
|
|
|
|
@torch.no_grad()
|
|
def reuse_packed_weight_storage(current: Any, replacement: Any) -> Any:
|
|
"""Reuse packed tensor addresses when a compatible weight is reloaded."""
|
|
if current is None or not _same_packed_layout(current, replacement):
|
|
return replacement
|
|
_copy_packed_tensors(current, replacement)
|
|
return current
|