[Mini-vLLM] SM86-only cut: models, backends, Rust, tests
- Keep only Qwen3.5, Qwen3.5-MTP and Qwen4Exp model paths; trim the model registry to 41 entries - Delete idle attention backends (MLA, FlashInfer, ROCm-only, quant backends) and dead GDN CuTeDSL kernels - Remove inc/quark quant loaders; keep compressed-tensors and Marlin - Drop the Rust frontend, pooling entrypoints, and idle test suites - Rebuild kernels for sm_80;sm_86 and strip the extensions - Verified: full import sweep clean (optional deps only) and a dummy-load Qwen3.5 engine run on RTX 3090 via the Triton GDN path Co-authored-by: Hinata <hinata@idihore.id> Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
@@ -26,14 +26,6 @@ from vllm.model_executor.warmup.flashinfer_autotune_cache import (
|
||||
resolve_flashinfer_autotune_v2_root,
|
||||
use_flashinfer_autotune_v2,
|
||||
)
|
||||
from vllm.model_executor.warmup.flashinfer_sparse_mla_warmup import (
|
||||
autotune_hisparse_flashinfer_attention,
|
||||
deepseek_v4_sparse_mla_attention_warmup,
|
||||
flashinfer_sparse_mla_decode_autotune_warmup,
|
||||
)
|
||||
from vllm.model_executor.warmup.kimi_k3_triton_warmup import (
|
||||
kimi_k3_triton_warmup,
|
||||
)
|
||||
from vllm.model_executor.warmup.mamba_triton_warmup import mamba_triton_warmup
|
||||
from vllm.model_executor.warmup.qwen4_exp_qsa_warmup import (
|
||||
qwen4_exp_qsa_triton_warmup,
|
||||
@@ -162,10 +154,6 @@ def _autotune_kimi_k3_kda_qkvg(model: torch.nn.Module) -> None:
|
||||
|
||||
|
||||
def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
from vllm.model_executor.warmup.minimax_m3_msa_warmup import (
|
||||
minimax_m3_msa_warmup,
|
||||
)
|
||||
|
||||
if not worker.use_v2_model_runner:
|
||||
# The KV-block zeroing kernel is driven by the scheduler's
|
||||
# `new_block_ids_to_zero`, so no dummy run ever reaches it.
|
||||
@@ -202,7 +190,6 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
|
||||
# Run next so input-prep kernels JIT against pristine runner state.
|
||||
if enable_jit_warmup:
|
||||
kimi_k3_triton_warmup(worker)
|
||||
watermark_sample_warmup(worker)
|
||||
qwen4_exp_qsa_triton_warmup(worker)
|
||||
|
||||
@@ -233,9 +220,6 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
|
||||
rocm_segmented_attn_autotune_warmup(worker)
|
||||
|
||||
flashinfer_sparse_mla_decode_autotune_warmup(worker)
|
||||
deepseek_v4_sparse_mla_attention_warmup(worker)
|
||||
|
||||
# Deep GEMM warmup
|
||||
do_deep_gemm_warmup = (
|
||||
is_deep_gemm_supported() and envs.VLLM_DEEP_GEMM_WARMUP != "skip"
|
||||
@@ -247,8 +231,6 @@ def kernel_warmup(worker: "Worker", *, process_local_only: bool = False):
|
||||
|
||||
b12x_warmup(worker, cudagraph_capture_sizes)
|
||||
|
||||
minimax_m3_msa_warmup(worker)
|
||||
|
||||
# Allocate the exact decode-sized workspace, autotune cache misses, and
|
||||
# resolve every CUDA Graph bucket before capture begins.
|
||||
# Lazy import: flashinfer_pcie_ipc_all_reduce imports flashinfer.comm,
|
||||
@@ -498,6 +480,10 @@ def flashinfer_autotune(runner: "GPUModelRunner", *, skip_attn: bool = False) ->
|
||||
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)
|
||||
_run_flashinfer_autotune_dummy_runs(runner, skip_attn=skip_attn)
|
||||
replayssm_autotune_warmup(runner)
|
||||
|
||||
Reference in New Issue
Block a user