[Model][NVIDIA] Route DSA models to the CUDA non-compiled path (#52861)

This commit is contained in:
Woosuk Kwon
2026-08-19 00:13:52 -07:00
committed by GitHub
parent ee11730751
commit b09bd69b5b
16 changed files with 260 additions and 146 deletions
+7 -5
View File
@@ -341,6 +341,7 @@ class SpeculativeConfig:
@staticmethod
def hf_config_override(hf_config: PretrainedConfig) -> PretrainedConfig:
initial_architecture = hf_config.architectures[0]
use_v32_mtp = hf_config.model_type in ("deepseek_v32", "glm_moe_dsa")
if hf_config.model_type == "dots3_note":
n_predict = getattr(hf_config, "num_nextn_predict_layers", 1)
mtp_layer_types = getattr(hf_config, "mtp_layer_types", None)
@@ -365,7 +366,12 @@ class SpeculativeConfig:
if hf_config.model_type == "deepseek_mtp":
n_predict = getattr(hf_config, "num_nextn_predict_layers", None)
hf_config.update(
{"n_predict": n_predict, "architectures": ["DeepSeekMTPModel"]}
{
"n_predict": n_predict,
"architectures": [
"DeepseekV32MTPModel" if use_v32_mtp else "DeepSeekMTPModel"
],
}
)
if hf_config.model_type == "deepseek_v4":
hf_config.model_type = "deepseek_mtp"
@@ -759,10 +765,6 @@ class SpeculativeConfig:
if self.method == "mtp":
if self.target_model_config is None:
raise ValueError("target_model_config must be present for mtp")
if self.target_model_config.hf_text_config.model_type == "deepseek_v32":
# FIXME(luccafong): cudagraph with v32 MTP is not supported,
# remove this when the issue is fixed.
self.enforce_eager = True
# use the draft model from the same model:
self.model = self.target_model_config.model
# Align the quantization of draft model for cases such as
+67 -39
View File
@@ -69,7 +69,9 @@ logger = init_logger(__name__)
DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
{
"DeepseekV2ForCausalLM",
"DeepseekV32ForCausalLM",
"DeepseekV4ForCausalLM",
"GlmMoeDsaForCausalLM",
"GraniteMoeForCausalLM",
"InklingForCausalLM",
"InklingForConditionalGeneration",
@@ -79,6 +81,23 @@ DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES = frozenset(
}
)
DEFAULT_BREAKABLE_CUDAGRAPH_ARCHITECTURES = frozenset(
{
"DeepseekV32MTPModel",
"DeepseekV32ForCausalLM",
"DeepseekV4ForCausalLM",
"DeepSeekV4MTPModel",
"GlmMoeDsaForCausalLM",
"InklingForCausalLM",
"InklingForConditionalGeneration",
"KimiK3ForConditionalGeneration",
"KimiK3MTPModel",
"KimiLinearForCausalLM",
"MiniMaxM3SparseForCausalLM",
"MiniMaxM3SparseForConditionalGeneration",
}
)
@lru_cache
def default_v2_model_runner_architectures() -> frozenset[str]:
@@ -86,13 +105,28 @@ def default_v2_model_runner_architectures() -> frozenset[str]:
from vllm.platforms import current_platform
if current_platform.is_rocm():
# TODO(rocm): DeepSeek V4 is still faster on MRV1 on ROCm. The
# attention layer picks the eager cudagraph region MRV1 needs, so
# this is a perf default only; drop it once MRV2 catches up.
return DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES - {"DeepseekV4ForCausalLM"}
# TODO(rocm): These models are either unsupported by MRV2 or slower with
# MRV2 on AMD GPUs.
return DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES - {
"DeepseekV32ForCausalLM",
"DeepseekV4ForCausalLM",
}
return DEFAULT_V2_MODEL_RUNNER_ARCHITECTURES
@lru_cache
def default_breakable_cudagraph_architectures() -> frozenset[str]:
"""Architectures defaulting to breakable CUDA graphs on this platform."""
from vllm.platforms import current_platform
if current_platform.is_rocm():
return DEFAULT_BREAKABLE_CUDAGRAPH_ARCHITECTURES - {
"DeepseekV32ForCausalLM",
"DeepseekV32MTPModel",
}
return DEFAULT_BREAKABLE_CUDAGRAPH_ARCHITECTURES
class OptimizationLevel(IntEnum):
"""Optimization level enum."""
@@ -691,6 +725,34 @@ class VllmConfig:
return False
return is_default_v2_architecture or not model_config.is_moe
def _uses_breakable_cudagraph_by_default(self) -> bool:
model_config = self.model_config
if model_config is None:
return False
architectures = set(model_config.architectures)
return bool(architectures & default_breakable_cudagraph_architectures())
def _maybe_enable_breakable_cudagraph(self) -> bool:
if (
"VLLM_USE_BREAKABLE_CUDAGRAPH" not in os.environ
and self._uses_breakable_cudagraph_by_default()
):
os.environ["VLLM_USE_BREAKABLE_CUDAGRAPH"] = "1"
logger.info_once(
"Auto-enabling VLLM_USE_BREAKABLE_CUDAGRAPH=1. "
"Set VLLM_USE_BREAKABLE_CUDAGRAPH=0 to opt out."
)
from vllm.compilation.breakable_cudagraph import (
is_breakable_cudagraph_enabled,
)
enabled = is_breakable_cudagraph_enabled()
if enabled:
self.compilation_config.mode = CompilationMode.NONE
return enabled
@property
def needs_dp_coordinator(self) -> bool:
"""
@@ -1288,41 +1350,7 @@ class VllmConfig:
)
self.compilation_config.mode = CompilationMode.NONE
# For model classes don't carry @support_torch_compile —
# the breakable cudagraph is the supported PIECEWISE path. Auto-enable
# it unless the user has explicitly opted out via the env var.
if (
self.model_config is not None
and "VLLM_USE_BREAKABLE_CUDAGRAPH" not in os.environ
and any(
a
in (
"DeepseekV4ForCausalLM",
"DeepSeekV4MTPModel",
"InklingForCausalLM",
"InklingForConditionalGeneration",
"KimiK3ForConditionalGeneration",
"KimiK3MTPModel",
"KimiLinearForCausalLM",
"MiniMaxM3SparseForCausalLM",
"MiniMaxM3SparseForConditionalGeneration",
)
for a in self.model_config.architectures
)
):
os.environ["VLLM_USE_BREAKABLE_CUDAGRAPH"] = "1"
logger.info_once(
"Auto-enabling VLLM_USE_BREAKABLE_CUDAGRAPH=1. "
"Set VLLM_USE_BREAKABLE_CUDAGRAPH=0 to opt out."
)
from vllm.compilation.breakable_cudagraph import (
is_breakable_cudagraph_enabled,
)
breakable_cudagraph_enabled = is_breakable_cudagraph_enabled()
if breakable_cudagraph_enabled:
self.compilation_config.mode = CompilationMode.NONE
breakable_cudagraph_enabled = self._maybe_enable_breakable_cudagraph()
if not breakable_cudagraph_enabled and (
self.compilation_config.backend == "eager"
+6 -2
View File
@@ -92,7 +92,10 @@ _TEXT_GENERATION_MODELS = {
"DeepseekForCausalLM": ("deepseek_v2", "DeepseekForCausalLM"),
"DeepseekV2ForCausalLM": ("deepseek_v2", "DeepseekV2ForCausalLM"),
"DeepseekV3ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
"DeepseekV32ForCausalLM": ("deepseek_v2", "DeepseekV3ForCausalLM"),
"DeepseekV32ForCausalLM": (
"vllm.models.deepseek_v32",
"DeepseekV32ForCausalLM",
),
"DeepseekV4ForCausalLM": ("vllm.models.deepseek_v4", "DeepseekV4ForCausalLM"),
"Ernie4_5ForCausalLM": ("ernie45", "Ernie4_5ForCausalLM"),
"Ernie4_5_MoeForCausalLM": ("ernie45_moe", "Ernie4_5_MoeForCausalLM"),
@@ -115,7 +118,7 @@ _TEXT_GENERATION_MODELS = {
"Glm4ForCausalLM": ("glm4", "Glm4ForCausalLM"),
"Glm4MoeForCausalLM": ("glm4_moe", "Glm4MoeForCausalLM"),
"Glm4MoeLiteForCausalLM": ("glm4_moe_lite", "Glm4MoeLiteForCausalLM"),
"GlmMoeDsaForCausalLM": ("deepseek_v2", "GlmMoeDsaForCausalLM"),
"GlmMoeDsaForCausalLM": ("vllm.models.deepseek_v32", "GlmMoeDsaForCausalLM"),
"GptOssForCausalLM": ("gpt_oss", "GptOssForCausalLM"),
"GPT2LMHeadModel": ("gpt2", "GPT2LMHeadModel"),
"GPTJForCausalLM": ("gpt_j", "GPTJForCausalLM"),
@@ -661,6 +664,7 @@ _SPECULATIVE_DECODING_MODELS = {
"Eagle3DeepseekV3ForCausalLM": ("deepseek_eagle3", "Eagle3DeepseekV2ForCausalLM"),
"EagleDeepSeekMTPModel": ("deepseek_eagle", "EagleDeepseekV3ForCausalLM"),
"DeepSeekMTPModel": ("deepseek_mtp", "DeepSeekMTP"),
"DeepseekV32MTPModel": ("vllm.models.deepseek_v32", "DeepseekV32MTP"),
"Dots3NoteMTPModel": ("vllm.models.dots3_note", "Dots3NoteMTP"),
"DeepSeekV4MTPModel": ("vllm.models.deepseek_v4", "DeepSeekV4MTP"),
"BailingMoeV3MTPModel": ("bailing_moe_v3_mtp", "BailingMoeV3MTPModel"),
+17 -8
View File
@@ -1,26 +1,35 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""DeepSeek V3.2 (``deepseek_v32``) model — hardware-isolated entry point.
"""DeepSeek V3.2 (``deepseek_v32``) platform entry point.
DeepSeek V3.2 introduced the DeepSeek Sparse Attention (DSA) architecture:
MLA + a "lightning indexer" that selects the top-k tokens for a sparse MLA
attend. The same model code serves any DSA checkpoint, including GLM-5.2
(``glm_moe_dsa``), which reuses this architecture.
The CUDA implementation selects capability-specific kernels internally and
falls back when an optimization is unavailable. Other platforms use the
generic implementation by default.
"""
from vllm.platforms import current_platform
if current_platform.is_rocm():
from .amd.model import DeepseekV32ForCausalLM
from .amd.mtp import DeepseekV32MTP
elif current_platform.is_xpu():
raise NotImplementedError("deepseek_v32 does not yet support XPU.")
else:
# Covers Blackwell (sm100) and all other CUDA devices.
if current_platform.is_cuda():
# GLM-5.2 (glm_moe_dsa) reuses the CUDA DSA module. Individual optimized
# kernels remain gated on the device capabilities they support.
from .nvidia.model import DeepseekV32ForCausalLM
from .nvidia.model import DeepseekV32ForCausalLM as GlmMoeDsaForCausalLM
from .nvidia.mtp import DeepseekV32MTP
else:
# ROCm, XPU, and CPU keep the generic implementation.
from vllm.model_executor.models.deepseek_mtp import DeepSeekMTP as DeepseekV32MTP
from vllm.model_executor.models.deepseek_v2 import (
DeepseekV3ForCausalLM as DeepseekV32ForCausalLM,
)
from vllm.model_executor.models.deepseek_v2 import GlmMoeDsaForCausalLM
__all__ = [
"DeepseekV32ForCausalLM",
"DeepseekV32MTP",
"GlmMoeDsaForCausalLM",
]
-1
View File
@@ -38,7 +38,6 @@ class DeepseekV32ROCmIndexer(DeepseekV32Indexer):
class DeepseekV32MLAAttention(DeepseekV32Attention):
require_fp8_kv_cache: bool = False
indexer_cls = DeepseekV32ROCmIndexer
def __init__(self, vllm_config, config, prefix, topk_indices_buffer=None):
+3 -16
View File
@@ -170,7 +170,6 @@ class DeepseekV32Attention(MLAAttention):
indexer: "DeepseekV32Indexer | None"
indexer_cls: "type[DeepseekV32Indexer]" = DeepseekV32Indexer
require_fp8_kv_cache: bool = True
supports_dense_mha_prefill = False
def __init__(
@@ -283,21 +282,9 @@ class DeepseekV32Attention(MLAAttention):
self.layer_name if enable_short_prefill_scoring_skip else ""
)
if self.require_fp8_kv_cache:
assert is_quantized_kv_cache(self.kv_cache_dtype), (
"deepseek_v32 (nvidia) requires an fp8 KV cache served by a sparse "
"MLA backend. Launch with --kv-cache-dtype fp8 (FlashInfer sparse) "
"or --kv-cache-dtype fp8_ds_mla (FlashMLA sparse)."
)
self._fp8_query = self.impl.supports_quant_query_input
if not self._fp8_query:
assert self.kv_cache_dtype == "fp8_ds_mla", (
"deepseek_v32 (nvidia) on a bf16-query sparse MLA backend "
"(FlashMLA sparse) requires the fp8_ds_mla KV cache layout. "
"Launch with --kv-cache-dtype fp8_ds_mla."
)
self._fp8_kv_needs_view = self.kv_cache_dtype != "fp8_ds_mla"
fp8_attention = is_quantized_kv_cache(self.kv_cache_dtype)
self._fp8_query = fp8_attention and self.impl.supports_quant_query_input
self._fp8_kv_needs_view = fp8_attention and self.kv_cache_dtype != "fp8_ds_mla"
self._index_rope_interleave = getattr(config, "indexer_rope_interleave", False)
+2 -1
View File
@@ -6,6 +6,7 @@ import torch
from vllm.platforms import current_platform
from vllm.triton_utils import tl, triton
from vllm.utils.torch_utils import is_quantized_kv_cache
# Cache of tiny 1-element dummy tensors (per device, dtype) reused by the
# has_indexer=False path so the indexer args don't allocate every call.
@@ -473,7 +474,7 @@ def fused_norm_rope(
# --- MLA KV cache setup ---
mla_cache_ds_mla = mla_kv_cache_dtype == "fp8_ds_mla"
mla_cache_fp8 = mla_kv_cache_dtype not in ("auto", "fp8_ds_mla")
mla_cache_fp8 = is_quantized_kv_cache(mla_kv_cache_dtype) and not mla_cache_ds_mla
mla_num_tiles = 1
mla_ds_scale_view = torch.empty(0, dtype=torch.float32, device=device)
mla_ds_rope_view = torch.empty(0, dtype=torch.bfloat16, device=device)
+5 -1
View File
@@ -1018,7 +1018,11 @@ class SpecDecodeBaseProposer:
# feedback into the next draft step.
architectures = self.draft_model_config.hf_config.architectures or []
return bool(
{"DeepSeekMTPModel", "KimiK3MTPModel"}.intersection(architectures)
{
"DeepSeekMTPModel",
"DeepseekV32MTPModel",
"KimiK3MTPModel",
}.intersection(architectures)
)
return self.method not in ("mtp", "draft_model", "dflash")