MOVZXandHinata
61fbe0b865
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>
98 lines
3.3 KiB
Python
98 lines
3.3 KiB
Python
# SPDX-License-Identifier: Apache-2.0
|
|
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
|
"""Helpers for reading and editing PyTorch's allocator configuration.
|
|
|
|
Torch's allocator config is parsed as a whole: ``parseArgs`` resets every
|
|
option that is not explicitly present in the string it is handed. Writing a
|
|
single field therefore silently drops the rest of the user's configuration,
|
|
for the remainder of the process:
|
|
|
|
>>> # PYTORCH_ALLOC_CONF=expandable_segments:True,max_split_size_mb:512
|
|
>>> torch._C._accelerator_setAllocatorSettings("expandable_segments:False")
|
|
>>> # max_split_size is now SIZE_MAX, not 512 MiB -- and writing
|
|
>>> # "expandable_segments:True" back does not bring it home either.
|
|
|
|
So anything that wants to toggle one option has to read the current config,
|
|
flip just that field, and write the whole string back.
|
|
"""
|
|
|
|
import os
|
|
|
|
import torch
|
|
|
|
# Checked in torch's own precedence order: the accelerator-agnostic variable
|
|
# wins, then the CUDA one, then the HIP one (which is what ROCm users set).
|
|
ALLOC_CONF_ENV_VARS = (
|
|
"PYTORCH_ALLOC_CONF",
|
|
"PYTORCH_CUDA_ALLOC_CONF",
|
|
"PYTORCH_HIP_ALLOC_CONF",
|
|
)
|
|
|
|
EXPANDABLE_SEGMENTS = "expandable_segments"
|
|
|
|
|
|
def alloc_conf_from_env(
|
|
environ: "os._Environ[str] | dict[str, str] | None" = None,
|
|
) -> str:
|
|
"""The allocator config string the process was started with, if any."""
|
|
env = os.environ if environ is None else environ
|
|
for name in ALLOC_CONF_ENV_VARS:
|
|
conf = env.get(name)
|
|
if conf:
|
|
return conf
|
|
return ""
|
|
|
|
|
|
def live_alloc_conf() -> str | None:
|
|
"""The allocator config string currently in effect, or ``None``.
|
|
|
|
Read from the live allocator rather than the environment, so that a
|
|
*runtime* write by another component is visible.
|
|
"""
|
|
try:
|
|
settings = torch.cuda.memory._snapshot().get("allocator_settings")
|
|
if settings is not None:
|
|
for name in ALLOC_CONF_ENV_VARS:
|
|
conf = settings.get(name)
|
|
if conf is not None:
|
|
return conf
|
|
except Exception: # noqa: BLE001
|
|
pass
|
|
return None
|
|
|
|
|
|
def current_alloc_conf() -> str:
|
|
"""The live allocator config, falling back to the environment."""
|
|
live = live_alloc_conf()
|
|
return live if live is not None else alloc_conf_from_env()
|
|
|
|
|
|
def conf_flag_enabled(conf: str, key: str) -> bool:
|
|
"""Parse ``<key>:<bool>`` out of an allocator config string."""
|
|
for field in conf.split(","):
|
|
name, _, value = field.partition(":")
|
|
if name.strip() == key:
|
|
return value.strip().lower() in ("true", "1")
|
|
return False
|
|
|
|
|
|
def with_conf_flag(conf: str, key: str, enabled: bool) -> str:
|
|
"""Return ``conf`` with only ``key`` flipped, every other field verbatim."""
|
|
want = f"{key}:{'True' if enabled else 'False'}"
|
|
fields = [f for f in (f.strip() for f in conf.split(",")) if f]
|
|
for i, field in enumerate(fields):
|
|
if field.partition(":")[0].strip() == key:
|
|
fields[i] = want
|
|
return ",".join(fields)
|
|
fields.append(want)
|
|
return ",".join(fields)
|
|
|
|
|
|
def set_alloc_conf(conf: str) -> None:
|
|
"""Write a complete allocator config string."""
|
|
setter = getattr(torch._C, "_accelerator_setAllocatorSettings", None)
|
|
if setter is not None:
|
|
setter(conf)
|
|
return
|
|
torch.cuda.memory._set_allocator_settings(conf)
|