[Mini-VLLM] Cut dead subsystems across serving, connectors, parsers
- Remove P/D transfer connectors: nixl, nixl_chat, mooncake x2, hf3fs, moriio, hisparse, example and hidden-states connectors - Reduce weight_transfer to the base API used by model loading - Remove kv_offload tiering, simple_kv_offload, ec cpu/example, snapshot command, engine snapshot server - Remove transformers model fallback, dsv4 tokenizers, dead platforms cpu/tpu/xpu, flashmla third_party, dsa, megamoe/flydsl, dead attn ops, dead layers (short_conv, fused_embed_norm, indexer_topk, relu2 gemma_ar, lightning_attn, minimax_rms_norm), gemm_rs_ar - Keep MoE LoRA op (imported eagerly by punica_gpu); drop fp8 variant - Rebuild parser hubs around Qwen3 adapters only (36 dead adapters) - Remove dead tests with the code they covered Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -6,7 +6,6 @@ happen during model execution.
|
||||
"""
|
||||
|
||||
import json
|
||||
import sys
|
||||
import time
|
||||
from pathlib import Path
|
||||
from typing import TYPE_CHECKING
|
||||
@@ -134,25 +133,6 @@ def _warmup_bf16x3_router_gemm(
|
||||
logger.info_once("Warmed up BF16x3 router GEMM configs: %s.", configs)
|
||||
|
||||
|
||||
def _warmup_gemm_rs_ar() -> None:
|
||||
# Model construction (Kimi-K3, DeepSeek-V4.1) imports this module only
|
||||
# when GEMM-RS/AR is enabled and initializes its singleton before
|
||||
# kernel_warmup runs. Avoid importing it here so other models do not
|
||||
# compile the RS/AR variants.
|
||||
module = sys.modules.get("vllm.model_executor.kernels.linear.cute_dsl.gemm_rs_ar")
|
||||
if module is None:
|
||||
return
|
||||
compiled = module.warmup_gemm_rs_ar()
|
||||
if compiled:
|
||||
logger.info_once("Warmed up %d GEMM-RS/AR variants.", compiled)
|
||||
|
||||
|
||||
def _autotune_kimi_k3_kda_qkvg(model: torch.nn.Module) -> None:
|
||||
module = sys.modules.get("vllm.models.kimi_k3.nvidia.low_latency_gemm")
|
||||
if module is not None:
|
||||
module.autotune_kda_qkvg(model)
|
||||
|
||||
|
||||
def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
if not worker.use_v2_model_runner:
|
||||
# The KV-block zeroing kernel is driven by the scheduler's
|
||||
@@ -202,8 +182,6 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
if current_platform.has_device_capability(90):
|
||||
_warmup_ll_bf16_router_gemm(worker.get_model())
|
||||
|
||||
_warmup_gemm_rs_ar()
|
||||
|
||||
if worker.vllm_config.kernel_config.enable_cutedsl_warmup:
|
||||
# TODO(roberto): Remove after registered CuTeDSL warmups are migrated
|
||||
# to the shared JIT warmup infrastructure.
|
||||
@@ -478,16 +456,11 @@ def flashinfer_autotune(runner: "GPUModelRunner", *, skip_attn: bool = False) ->
|
||||
)
|
||||
skip_attn = skip_attn or hisparse_enabled
|
||||
if hisparse_enabled:
|
||||
# HiSparse hot-buffer attention is bounded by decode batch
|
||||
# size, not the prefill-sized batch used for the full model.
|
||||
from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import (
|
||||
autotune_hisparse_flashinfer_attention,
|
||||
)
|
||||
|
||||
autotune_hisparse_flashinfer_attention(runner)
|
||||
# HiSparse config still bounds decode-batch autotune even
|
||||
# though the connector was removed from this build.
|
||||
logger.info_once("HiSparse configured; skipping attention autotune.")
|
||||
_run_flashinfer_autotune_dummy_runs(runner, skip_attn=skip_attn)
|
||||
replayssm_autotune_warmup(runner)
|
||||
_autotune_kimi_k3_kda_qkvg(runner.get_model())
|
||||
finally:
|
||||
set_autotune_process_group(None)
|
||||
|
||||
|
||||
Reference in New Issue
Block a user