[Model][ROCm] Enable DeepSeek V4 Vision (#55107)

Signed-off-by: Andreas Karatzas <Andreas.Karatzas@amd.com>
Signed-off-by: Andreas Karatzas <akaratza@amd.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Andreas Karatzas
2026-09-11 09:17:30 -07:00
committed by GitHub
co-authored by OpenAI Codex
parent 127143e27f
commit 9dd969da09
17 changed files with 1055 additions and 347 deletions
+4 -2
View File
@@ -48,8 +48,9 @@ steps:
- tests/models/test_terratorch.py
- tests/models/transformers/test_backend.py
- tests/models/test_registry.py
- tests/models/test_deepseek_v4_vl_rocm.py
commands:
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py
mirror:
amd:
label: ":amd: (MI300) Basic Models (Other)"
@@ -59,9 +60,10 @@ steps:
depends_on:
- image-build-amd
source_file_dependencies:
- tests/models/test_deepseek_v4_vl_rocm.py
- tests/models/test_hyv4_rocm.py
commands:
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py
- pytest -v -s models/test_terratorch.py models/transformers/test_backend.py models/test_registry.py models/test_deepseek_v4_vl_rocm.py
- VLLM_ROCM_USE_AITER=1 pytest -v -s models/test_hyv4_rocm.py -m 'not distributed'
- label: ":nvidia: (B200) Inkling"
@@ -336,6 +336,101 @@ def test_compute_global_topk_ragged_indices_and_indptr() -> None:
torch.testing.assert_close(actual_lens, expected_lens)
@torch.inference_mode()
def test_combine_topk_swa_indices_adds_image_visibility() -> None:
from vllm.models.deepseek_v4.amd.rocm import combine_topk_swa_indices
device = torch.device("cuda")
num_tokens = 8
topk_indices = torch.full((num_tokens, 1), -1, dtype=torch.int32, device=device)
query_start_loc = torch.tensor([0, num_tokens], dtype=torch.int32, device=device)
seq_lens = torch.tensor([num_tokens], dtype=torch.int32, device=device)
gather_lens = seq_lens.clone()
left_visible = torch.tensor(
[0, 0, 0, 1, 2, 3, 4, 0], dtype=torch.int32, device=device
)
right_visible = torch.tensor(
[0, 0, 4, 3, 2, 1, 0, 0], dtype=torch.int32, device=device
)
indices, lens = combine_topk_swa_indices(
topk_indices,
query_start_loc,
seq_lens,
gather_lens,
window_size=4,
compress_ratio=1,
topk=0,
M=16,
N=0,
max_image_tokens=5,
left_visible=left_visible,
right_visible=right_visible,
)
expected_rows = [
[0],
[0, 1],
[0, 1, 2, 3, 4, 5, 6],
[0, 1, 2, 3, 4, 5, 6],
[1, 2, 3, 4, 5, 6],
[2, 3, 4, 5, 6],
[2, 3, 4, 5, 6],
[4, 5, 6, 7],
]
for token_idx, expected in enumerate(expected_rows):
actual = indices[token_idx, : lens[token_idx]].cpu().tolist()
assert actual == expected
@torch.inference_mode()
def test_combine_topk_swa_indices_apc_hit_inside_image() -> None:
from vllm.models.deepseek_v4.amd.rocm import combine_topk_swa_indices
device = torch.device("cuda")
indices, lens = combine_topk_swa_indices(
torch.full((2, 1), -1, dtype=torch.int32, device=device),
torch.tensor([0, 2], dtype=torch.int32, device=device),
torch.tensor([10], dtype=torch.int32, device=device),
# Only positions [5, 10) exist in the gathered SWA workspace.
torch.tensor([5], dtype=torch.int32, device=device),
window_size=4,
compress_ratio=1,
topk=0,
M=10,
N=0,
max_image_tokens=10,
left_visible=torch.tensor([8, 9], dtype=torch.int32, device=device),
# Deliberately extends beyond seq_len to exercise the upper clamp too.
right_visible=torch.tensor([5, 5], dtype=torch.int32, device=device),
)
assert lens.cpu().tolist() == [5, 5]
assert indices[:, :5].cpu().tolist() == [list(range(5)), list(range(5))]
@torch.inference_mode()
def test_combine_topk_swa_indices_keeps_vision_row_width_without_images() -> None:
from vllm.models.deepseek_v4.amd.rocm import combine_topk_swa_indices
device = torch.device("cuda")
indices, lens = combine_topk_swa_indices(
torch.full((1, 1), -1, dtype=torch.int32, device=device),
torch.tensor([0, 1], dtype=torch.int32, device=device),
torch.tensor([1], dtype=torch.int32, device=device),
torch.tensor([1], dtype=torch.int32, device=device),
window_size=120,
compress_ratio=1,
topk=0,
M=256,
N=0,
max_image_tokens=16,
)
assert indices.shape == (1, 256)
assert lens.item() == 1
def test_extra_cache_nan_free_provenance_gate(monkeypatch) -> None:
from vllm.models.deepseek_v4.amd import rocm as mod
@@ -164,10 +164,11 @@ def test_model_tensor_schema(model_id: str):
"Kimi-K2.5's offline inference has issues about vision chunks. Fix later."
)
if model_id == "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp" and not (
current_platform.is_cuda()
if (
model_id == "deepseek-ai/DeepSeek-V4-Flash-Vision-Exp"
and not current_platform.is_cuda_alike()
):
pytest.skip("Deepseek V4 is only supported on CUDA")
pytest.skip("Deepseek V4 vision is only supported on CUDA and ROCm")
if model_id == "zai-org/GLM-5.3-Flash" and (current_platform.is_xpu()):
pytest.skip("GLM-5.3-Flash is not supported on XPU")
+458
View File
@@ -0,0 +1,458 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from types import SimpleNamespace
import pytest
import torch
from torch import nn
from vllm.model_executor.models.utils import WeightsMapper
from vllm.platforms import current_platform
pytestmark = pytest.mark.skipif(
not current_platform.is_rocm(), reason="ROCm-specific tests"
)
def _preshuffled_fp8_linear(
holder: str = "quant_method", weight_shape: tuple[int, int] = (256, 128)
) -> nn.Module:
pytest.importorskip("aiter")
from vllm.model_executor.kernels.linear.scaled_mm.aiter import (
AiterPreshuffledFp8BlockScaledMMKernel,
)
from vllm.model_executor.kernels.linear.scaled_mm.ScaledMMLinearKernel import (
FP8ScaledMMLinearLayerConfig,
)
from vllm.model_executor.layers.quantization.utils.quant_utils import (
kFp8Dynamic128Sym,
kFp8Static128BlockSym,
)
config = FP8ScaledMMLinearLayerConfig(
weight_quant_key=kFp8Static128BlockSym,
activation_quant_key=kFp8Dynamic128Sym,
input_dtype=torch.bfloat16,
out_dtype=torch.bfloat16,
weight_shape=weight_shape,
)
layer = nn.Module()
layer.weight = nn.Parameter(
torch.randn(*weight_shape, device="cuda").to(current_platform.fp8_dtype()),
requires_grad=False,
)
layer.register_parameter(
"weight_scale" if holder == "scheme" else "weight_scale_inv",
nn.Parameter(
torch.ones(*(dim // 128 for dim in weight_shape), device="cuda"),
requires_grad=False,
),
)
setattr(
layer,
holder,
SimpleNamespace(fp8_linear=AiterPreshuffledFp8BlockScaledMMKernel(config)),
)
return layer
@pytest.mark.parametrize("holder", ["quant_method", "scheme"])
def test_rocm_wo_a_keeps_row_major_weights(holder: str, default_vllm_config) -> None:
from vllm.v1.attention.ops.rocm_aiter_mla_sparse import _get_cached_wo_a_bf16
layer = _preshuffled_fp8_linear(holder)
layer.is_bmm = True
original = layer.weight.detach().clone()
kernel = getattr(layer, holder).fp8_linear
kernel.process_weights_after_loading(layer)
actual = _get_cached_wo_a_bf16(layer, 2, 128, 128)
torch.testing.assert_close(actual, original.to(torch.bfloat16).view(2, 128, 128))
@pytest.mark.parametrize("holder", ["quant_method", "scheme"])
def test_rocm_gateup_shuffles_once_and_down_proj_keeps_preshuffled_backend(
holder: str, monkeypatch: pytest.MonkeyPatch, default_vllm_config
) -> None:
from vllm._aiter_ops import rocm_aiter_ops
from vllm.models.deepseek_v4.amd import model as rocm_model
gate_up, down = _preshuffled_fp8_linear(holder), _preshuffled_fp8_linear(holder)
original_gate_up = gate_up.weight.detach().clone()
original_down = down.weight.detach().clone()
monkeypatch.setattr(
rocm_model, "MergedColumnParallelLinear", lambda *a, **k: gate_up
)
monkeypatch.setattr(rocm_model, "RowParallelLinear", lambda *a, **k: down)
monkeypatch.setattr(rocm_aiter_ops, "is_enabled", lambda: True)
model = rocm_model.DeepseekV4MLP(128, 128, "silu")
for linear in (gate_up, down):
getattr(linear, holder).fp8_linear.process_weights_after_loading(linear)
model.prepare_gateup_preshuffle()
for linear, original in ((gate_up, original_gate_up), (down, original_down)):
expected = rocm_aiter_ops.shuffle_weight(original, layout=(16, 16))
torch.testing.assert_close(
linear.weight.view(torch.uint8), expected.view(torch.uint8)
)
def test_rocm_fused_qkv_quant_matches_preshuffled_gemm_scale_layout(
monkeypatch: pytest.MonkeyPatch, default_vllm_config
) -> None:
from vllm._aiter_ops import rocm_aiter_ops
from vllm.models.deepseek_v4.amd.rocm import (
DeepseekV4ROCMAiterMLAAttention,
apply_pre_quantized_block_scaled_mm,
)
if not rocm_aiter_ops.is_blockscale_bpreshuffle_tuned(1024, 256):
pytest.skip("Requires a tuned AITER preshuffled FP8 GEMM")
linear = _preshuffled_fp8_linear(weight_shape=(1024, 256))
original = linear.weight.detach().clone()
linear.quant_method.fp8_linear.process_weights_after_loading(linear)
attention = DeepseekV4ROCMAiterMLAAttention.__new__(DeepseekV4ROCMAiterMLAAttention)
nn.Module.__init__(attention)
attention.wq_b = linear
attention.indexer = None
attention.q_lora_rank, attention.head_dim = 256, 128
attention.eps = 1e-5
attention.q_norm = SimpleNamespace(
weight=torch.ones(256, dtype=torch.bfloat16, device="cuda")
)
attention.kv_norm = SimpleNamespace(
weight=torch.ones(128, dtype=torch.bfloat16, device="cuda")
)
monkeypatch.setattr(rocm_aiter_ops, "is_linear_fp8_enabled", lambda: True)
# Multiple rows and groups with different ranges expose byte-order errors.
inputs = torch.randn(16, 384, dtype=torch.bfloat16, device="cuda")
inputs[:, :128] *= 4
quantized, scales, _ = attention._split_qkv_and_norm(inputs)
assert scales is not None
output = apply_pre_quantized_block_scaled_mm(linear, quantized, scales)
q = inputs[:, :256].float()
normalized = q * torch.rsqrt(q.square().mean(dim=-1, keepdim=True) + attention.eps)
reference_scales = normalized.view(16, 2, 128).abs().amax(dim=-1)
reference_scales /= torch.finfo(current_platform.fp8_dtype()).max
dequantized = quantized.float() * reference_scales.repeat_interleave(128, dim=1)
expected = dequantized @ original.float().T
torch.testing.assert_close(output.float(), expected, atol=0.125, rtol=0.01)
def test_rocm_packed_kv_cache_auto_uses_ds_mla_layout() -> None:
from vllm.config import CacheConfig
from vllm.models.deepseek_v4.attention import _resolve_dsv4_kv_cache_dtype
cache_config = CacheConfig()
resolved_dtype, torch_dtype = _resolve_dsv4_kv_cache_dtype(
use_fp8_ds_mla_layout=True,
kv_cache_dtype=cache_config.cache_dtype,
cache_config=cache_config,
)
assert resolved_dtype == "fp8_ds_mla"
assert torch_dtype is torch.uint8
assert cache_config.cache_dtype == "fp8_ds_mla"
def test_rocm_packed_kv_cache_rejects_unquantized_dtype() -> None:
from vllm.config import CacheConfig
from vllm.models.deepseek_v4.attention import _resolve_dsv4_kv_cache_dtype
cache_config = CacheConfig(cache_dtype="bfloat16")
with pytest.raises(ValueError, match="only supports fp8 kv-cache"):
_resolve_dsv4_kv_cache_dtype(
use_fp8_ds_mla_layout=True,
kv_cache_dtype=cache_config.cache_dtype,
cache_config=cache_config,
)
def test_vl_mapper_preserves_rocm_weight_mapping() -> None:
from vllm.models.deepseek_v4.amd.model import _make_deepseek_v4_weights_mapper
from vllm.models.deepseek_v4.common.vl_model import (
_make_deepseek_v4_vl_weights_mapper,
)
text_mapper = _make_deepseek_v4_weights_mapper("fp4", fuse_shared_experts=True)
mapper = _make_deepseek_v4_vl_weights_mapper(text_mapper, image_enabled=True)
assert mapper._map_name("layers.3.attn.wq_a.input_scale") == (
"language_model.model.layers.3.attn.wq_a.input_scale_2"
)
assert mapper._map_name("layers.3.ffn.shared_experts.w2.weight") == (
"language_model.model.layers.3.ffn.shared_experts.w2.weight"
)
assert mapper._map_name("head.weight") == "language_model.lm_head.weight"
def test_rocm_moe_wires_vision_routing_on_hash_and_regular_layers(
monkeypatch: pytest.MonkeyPatch,
) -> None:
from vllm.models.deepseek_v4.amd import model as rocm_model
from vllm.models.deepseek_v4.common.mm_preprocess import IMAGE_SENTINEL_BASE_ID
captured: list[dict] = []
class FakeGate(nn.Module):
def __init__(self, **kwargs) -> None:
super().__init__()
def fake_factory(**kwargs):
captured.append(kwargs)
return nn.Identity()
monkeypatch.setattr(rocm_model, "GateLinear", FakeGate)
monkeypatch.setattr(rocm_model, "FusedMoEFactory", fake_factory)
monkeypatch.setattr(rocm_model, "get_tensor_model_parallel_world_size", lambda: 1)
monkeypatch.setattr(rocm_model, "get_tensor_model_parallel_rank", lambda: 0)
config = SimpleNamespace(
hidden_size=16,
n_routed_experts=4,
num_experts_per_tok=2,
moe_intermediate_size=8,
swiglu_limit=None,
norm_topk_prob=True,
scoring_func="sqrtsoftplus",
num_hash_layers=1,
vocab_size=32,
topk_method="noaux_tc",
vision_n_layers=1,
n_shared_experts=None,
)
vllm_config = SimpleNamespace(
model_config=SimpleNamespace(hf_config=config), quant_config=None
)
hash_moe = rocm_model.DeepseekV4MoE(vllm_config, prefix="model.layers.0.ffn")
regular_moe = rocm_model.DeepseekV4MoE(vllm_config, prefix="model.layers.1.ffn")
assert hash_moe.gate.tid2eid is not None
assert regular_moe.gate.tid2eid is None
for moe, factory_kwargs in zip((hash_moe, regular_moe), captured, strict=True):
assert moe.gate.e_score_correction_bias is not None
assert moe.gate.bias_vl is not None
assert factory_kwargs["bias_vl"] is moe.gate.bias_vl
assert factory_kwargs["image_sentinel_lo"] == IMAGE_SENTINEL_BASE_ID
def test_rocm_mtp_forwards_input_ids_for_vision_routing(
monkeypatch: pytest.MonkeyPatch,
) -> None:
from vllm.models.deepseek_v4.amd import mtp as rocm_mtp
hidden_size = 4
hc_mult = 2
class FakeNorm(nn.Module):
def __init__(self) -> None:
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = 1e-6
class FakeMTPBlock(nn.Module):
use_fused_mhc = False
def __init__(self) -> None:
super().__init__()
self.seen_input_ids: torch.Tensor | None = None
def forward(
self,
*,
positions: torch.Tensor,
x: torch.Tensor,
input_ids: torch.Tensor | None,
):
self.seen_input_ids = input_ids
return x, None, None, None
def passthrough_mtp_input(
inputs_embeds: torch.Tensor,
positions: torch.Tensor,
previous_hidden_states: torch.Tensor,
*args,
) -> tuple[torch.Tensor, torch.Tensor]:
return inputs_embeds, previous_hidden_states
monkeypatch.setattr(
rocm_mtp, "_FUSED_MTP_INPUT_RMSNORM_KERNEL", passthrough_mtp_input
)
layer = object.__new__(rocm_mtp.DeepSeekV4MultiTokenPredictorLayer)
nn.Module.__init__(layer)
layer.config = SimpleNamespace(hidden_size=hidden_size)
layer.hc_mult = hc_mult
layer.enorm = FakeNorm()
layer.hnorm = FakeNorm()
layer.e_proj = nn.Identity()
layer.h_proj = nn.Identity()
layer.mtp_block = FakeMTPBlock()
input_ids = torch.tensor([11, 12])
positions = torch.tensor([3, 4])
inputs_embeds = torch.arange(8, dtype=torch.float32).view(2, hidden_size)
previous_hidden_states = torch.arange(16, dtype=torch.float32).view(2, -1)
output = layer(
input_ids,
positions,
previous_hidden_states,
inputs_embeds,
)
assert layer.mtp_block.seen_input_ids is input_ids
expected = previous_hidden_states.view(2, hc_mult, hidden_size)
expected = expected + inputs_embeds.unsqueeze(-2)
torch.testing.assert_close(output, expected.flatten(1))
def test_rocm_compute_logits_local_skips_gather() -> None:
from vllm.models.deepseek_v4.amd.model import DeepseekV4ForCausalLM
calls: list[tuple[nn.Module, torch.Tensor, bool]] = []
def logits_processor(
lm_head: nn.Module, hidden_states: torch.Tensor, *, skip_gather: bool = False
) -> torch.Tensor:
calls.append((lm_head, hidden_states, skip_gather))
return hidden_states + 1
model = object.__new__(DeepseekV4ForCausalLM)
nn.Module.__init__(model)
model.lm_head = nn.Identity()
model.logits_processor = logits_processor
hidden_states = torch.tensor([4.0])
result = model.compute_logits_local(hidden_states)
assert torch.equal(result, torch.tensor([5.0]))
assert calls == [(model.lm_head, hidden_states, True)]
class _FakeLanguageModel(nn.Module):
finalizes_weights_during_load = False
def __init__(self) -> None:
super().__init__()
self.tensor_a = nn.Parameter(torch.zeros(1))
self.tensor_c = nn.Parameter(torch.zeros(1))
self.finalized_values: list[tuple[float, float]] = []
def process_weights_after_loading(self) -> None:
self.finalized_values.append((self.tensor_a.item(), self.tensor_c.item()))
def compute_logits_local(self, hidden_states: torch.Tensor) -> torch.Tensor:
return hidden_states + 1
def test_vl_wrapper_streams_then_delegates_finalization() -> None:
from vllm.models.deepseek_v4.common.vl_model import (
DeepseekV4ForConditionalGeneration,
)
model = object.__new__(DeepseekV4ForConditionalGeneration)
nn.Module.__init__(model)
model.language_model = _FakeLanguageModel()
model.vision = nn.Module()
model.vision.tensor_b = nn.Parameter(torch.zeros(1))
model.hf_to_vllm_mapper = WeightsMapper()
def interleaved_weights():
yield "language_model.tensor_a", torch.tensor([1.0])
assert model.language_model.tensor_a.item() == 1.0
yield "vision.tensor_b", torch.tensor([2.0])
assert model.vision.tensor_b.item() == 2.0
yield "language_model.tensor_c", torch.tensor([3.0])
loaded = model.load_weights(interleaved_weights())
assert loaded == {
"language_model.tensor_a",
"vision.tensor_b",
"language_model.tensor_c",
}
assert model.language_model.finalized_values == []
model.process_weights_after_loading()
assert model.language_model.finalized_values == [(1.0, 3.0)]
assert torch.equal(
model.compute_logits_local(torch.tensor([4.0])), torch.tensor([5.0])
)
model.process_weights_after_loading()
assert model.language_model.finalized_values == [(1.0, 3.0)]
class _FakeFinalizingLanguageModel(_FakeLanguageModel):
finalizes_weights_during_load = True
def __init__(self) -> None:
super().__init__()
self.load_calls = 0
def load_weights(self, weights) -> set[str]:
self.load_calls += 1
loaded = set()
for name, value in weights:
getattr(self, name).data.copy_(value)
loaded.add(name)
self.process_weights_after_loading()
return loaded
def test_vl_wrapper_groups_child_that_finalizes_during_load() -> None:
from vllm.models.deepseek_v4.common.vl_model import (
DeepseekV4ForConditionalGeneration,
)
model = object.__new__(DeepseekV4ForConditionalGeneration)
nn.Module.__init__(model)
model.language_model = _FakeFinalizingLanguageModel()
model.vision = nn.Module()
model.vision.tensor_b = nn.Parameter(torch.zeros(1))
model.hf_to_vllm_mapper = WeightsMapper()
loaded = model.load_weights(
iter(
(
("language_model.tensor_a", torch.tensor([1.0])),
("vision.tensor_b", torch.tensor([2.0])),
("language_model.tensor_c", torch.tensor([3.0])),
)
)
)
assert loaded == {
"language_model.tensor_a",
"vision.tensor_b",
"language_model.tensor_c",
}
assert model.language_model.load_calls == 1
assert model.language_model.finalized_values == [(1.0, 3.0)]
# The framework's later model-level hook must not double-finalize a child
# which already completed this work in load_weights.
model.process_weights_after_loading()
assert model.language_model.finalized_values == [(1.0, 3.0)]
def test_vl_wrapper_dummy_load_delegates_finalization() -> None:
from vllm.models.deepseek_v4.common.vl_model import (
DeepseekV4ForConditionalGeneration,
)
model = object.__new__(DeepseekV4ForConditionalGeneration)
nn.Module.__init__(model)
model.language_model = _FakeFinalizingLanguageModel()
# DummyModelLoader bypasses model.load_weights(), so no finalized marker
# exists and the framework-level hook must still delegate to the child.
model.process_weights_after_loading()
assert model.language_model.finalized_values == [(0.0, 0.0)]
model.process_weights_after_loading()
assert model.language_model.finalized_values == [(0.0, 0.0)]
+2 -2
View File
@@ -135,8 +135,8 @@ def can_initialize(
if model_arch == "DeepseekV4ForConditionalGeneration":
from vllm.platforms import current_platform
if not current_platform.is_cuda():
pytest.skip("Deepseek V4 is only supported on CUDA")
if not (current_platform.is_cuda() or current_platform.is_rocm()):
pytest.skip("Deepseek V4 vision is only supported on CUDA and ROCm")
with (
patch.object(V1EngineCore, "_initialize_kv_caches", _initialize_kv_caches_v1),
+4 -3
View File
@@ -64,10 +64,11 @@ def test_registry_imports(model_arch):
):
pytest.skip("HY V4 is only supported on CUDA and ROCm")
if model_arch == "DeepseekV4ForConditionalGeneration" and not (
current_platform.is_cuda()
if (
model_arch == "DeepseekV4ForConditionalGeneration"
and not current_platform.is_cuda_alike()
):
pytest.skip("Deepseek V4 is only supported on CUDA")
pytest.skip("Deepseek V4 vision is only supported on CUDA and ROCm")
# Ensure all model classes can be imported successfully
model_cls = ModelRegistry._try_load_model_cls(model_arch)
+39
View File
@@ -57,6 +57,45 @@ def _write_json(path: Path, value: object) -> None:
path.write_text(json.dumps(value), encoding="utf-8")
@pytest.mark.skipif(not current_platform.is_rocm(), reason="ROCm-specific test")
@pytest.mark.parametrize(
("is_mm_prefix_lm", "is_multimodal_model", "expected"),
[
pytest.param(True, True, True, id="multimodal-prefix-lm"),
pytest.param(False, True, False, id="multimodal-causal"),
pytest.param(True, False, False, id="text-prefix-lm"),
pytest.param(None, True, False, id="missing-model-config"),
],
)
def test_rocm_mm_prefix_lm_disables_chunked_mm_input(
is_mm_prefix_lm: bool | None,
is_multimodal_model: bool,
expected: bool,
) -> None:
from vllm.platforms.rocm import RocmPlatform
config = SimpleNamespace(
compilation_config=SimpleNamespace(cudagraph_mode=CUDAGraphMode.NONE),
parallel_config=SimpleNamespace(
prefill_context_parallel_size=1,
worker_cls="test-worker",
),
model_config=(
None
if is_mm_prefix_lm is None
else SimpleNamespace(is_mm_prefix_lm=is_mm_prefix_lm)
),
scheduler_config=SimpleNamespace(
is_multimodal_model=is_multimodal_model,
disable_chunked_mm_input=False,
),
)
RocmPlatform.check_and_update_config(config)
assert config.scheduler_config.disable_chunked_mm_input is expected
def test_kda_recoverssm_derivation_is_revalidated():
config = SimpleNamespace(
cache_config=SimpleNamespace(
@@ -42,6 +42,7 @@ def make_manager(
("mode", "piecewise_capture_available", "expected"),
[
("FULL_DECODE_ONLY", True, "FULL_DECODE_ONLY"),
("FULL_DECODE_ONLY", False, "FULL_DECODE_ONLY"),
("FULL", True, "FULL_AND_PIECEWISE"),
("FULL", False, "FULL_DECODE_ONLY"),
("FULL_AND_PIECEWISE", True, "FULL_AND_PIECEWISE"),
+4 -4
View File
@@ -20,7 +20,7 @@ if current_platform.is_rocm():
)
from .amd.model import DeepseekV4ForCausalLM
from .amd.mtp import DeepSeekV4MTP
from .vl_stub import ( # type: ignore[assignment]
from .common.vl_model import ( # type: ignore[assignment]
DeepseekV4ForConditionalGeneration,
)
elif current_platform.is_xpu():
@@ -38,14 +38,14 @@ elif current_platform.is_cpu():
DeepseekV4ForConditionalGeneration,
)
else:
from .common.vl_model import ( # type: ignore[assignment]
DeepseekV4ForConditionalGeneration,
)
from .nvidia.dspark import ( # type: ignore[assignment]
DSparkDeepseekV4ForCausalLM,
)
from .nvidia.model import DeepseekV4ForCausalLM # type: ignore[assignment]
from .nvidia.mtp import DeepSeekV4MTP # type: ignore[assignment]
from .nvidia.vl_model import ( # type: ignore[assignment]
DeepseekV4ForConditionalGeneration,
)
__all__ = [
"DSparkDeepseekV4ForCausalLM",
+23 -1
View File
@@ -79,6 +79,8 @@ from vllm.platforms import current_platform
from vllm.platforms.rocm import on_gfx950
from vllm.sequence import IntermediateTensors
from ..common.mm_preprocess import IMAGE_SENTINEL_BASE_ID
logger = init_logger(__name__)
@@ -549,6 +551,10 @@ class DeepseekV4MoE(nn.Module):
self.gate.e_score_correction_bias = None
self.gate.tid2eid = None
self.gate.bias_vl = None
self.image_sentinel_lo = (
IMAGE_SENTINEL_BASE_ID if getattr(config, "vision_n_layers", 0) > 0 else 0
)
is_hash_moe = extract_layer_index(prefix) < config.num_hash_layers
self.hash_indices_dtype = torch.int32
if is_hash_moe:
@@ -564,12 +570,20 @@ class DeepseekV4MoE(nn.Module):
),
requires_grad=False,
)
elif getattr(config, "topk_method", None) == "noaux_tc":
if getattr(config, "topk_method", None) == "noaux_tc" and (
not is_hash_moe or getattr(config, "vision_n_layers", 0) > 0
):
self.gate.e_score_correction_bias = nn.Parameter(
torch.empty(config.n_routed_experts, dtype=torch.float32),
requires_grad=False,
)
if getattr(config, "vision_n_layers", 0) > 0:
self.gate.bias_vl = nn.Parameter(
torch.empty(config.n_routed_experts, dtype=torch.float32),
requires_grad=False,
)
self.n_shared_experts = config.n_shared_experts
# TODO: Historically, only `VLLM_ROCM_USE_AITER_FUSION_SHARED_EXPERTS=1`
@@ -638,6 +652,8 @@ class DeepseekV4MoE(nn.Module):
routed_scaling_factor=self.routed_scaling_factor,
e_score_correction_bias=self.gate.e_score_correction_bias,
hash_indices_table=self.gate.tid2eid,
bias_vl=self.gate.bias_vl,
image_sentinel_lo=self.image_sentinel_lo,
swiglu_limit=self.swiglu_limit,
router_logits_dtype=torch.float32,
routed_experts_cls=(
@@ -660,6 +676,8 @@ class DeepseekV4MoE(nn.Module):
) -> torch.Tensor:
if self.gate.tid2eid is not None and input_ids is None:
raise ValueError("DeepSeek V4 hash MoE routing requires input_ids.")
if self.gate.bias_vl is not None and input_ids is None:
raise ValueError("DeepSeek V4 vision MoE routing requires input_ids.")
org_shape = hidden_states.shape
final_hidden_states = self.experts(
@@ -1351,6 +1369,7 @@ def _make_deepseek_v4_weights_mapper(
class DeepseekV4ForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
model_cls = DeepseekV4Model
finalizes_weights_during_load = False
# Default mapper assumes the original FP4-expert checkpoint layout.
# Overridden per-instance in __init__ when expert_dtype != "fp4".
@@ -1397,6 +1416,9 @@ class DeepseekV4ForCausalLM(nn.Module, SupportsPP, SupportsEagle3):
logits = self.logits_processor(self.lm_head, hidden_states)
return logits
def compute_logits_local(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.logits_processor(self.lm_head, hidden_states, skip_gather=True)
def forward(
self,
input_ids: torch.Tensor,
+1 -1
View File
@@ -160,7 +160,7 @@ class DeepSeekV4MultiTokenPredictorLayer(nn.Module):
inputs_embeds
).unsqueeze(-2)
hidden_states, residual, post_mix, res_mix = self.mtp_block(
positions=positions, x=hidden_states, input_ids=None
positions=positions, x=hidden_states, input_ids=input_ids
)
if self.mtp_block.use_fused_mhc:
hidden_states = self.mtp_block.hc_post(
+61 -7
View File
@@ -67,8 +67,13 @@ def weight_already_preshuffled(linear: torch.nn.Module) -> bool:
The hand-shuffles below (fused_wqa_wkv, wo_b, gate_up_proj) must be skipped
for those, since shuffle_weight is a permutation rather than an involution.
"""
kernel = getattr(getattr(linear, "quant_method", None), "fp8_linear", None)
return bool(getattr(kernel, "preshuffles_weight", False))
return any(
getattr(getattr(method, "fp8_linear", None), "preshuffles_weight", False)
for method in (
getattr(linear, "quant_method", None),
getattr(linear, "scheme", None),
)
)
def apply_pre_quantized_block_scaled_mm(
@@ -116,13 +121,18 @@ def _combine_topk_swa_indices_kernel(
query_start_loc_ptr,
seq_lens_ptr,
gather_lens_ptr,
left_visible_ptr,
right_visible_ptr,
M,
N,
TOP_K: tl.constexpr,
COMPRESS_RATIO: tl.constexpr,
WINDOW_SIZE: tl.constexpr,
SWA_WIDTH: tl.constexpr,
TOPK_WIDTH: tl.constexpr,
PADDED_TOP_K: tl.constexpr,
PADDED_SWA_WIDTH: tl.constexpr,
HAS_IMAGE: tl.constexpr,
):
batch_idx = tl.program_id(0)
worker_id = tl.program_id(1)
@@ -141,7 +151,20 @@ def _combine_topk_swa_indices_kernel(
token_idx_in_query = token_idx - query_start
pos = start_pos + token_idx_in_query
topk_len = tl.minimum((pos + 1) // COMPRESS_RATIO, TOP_K)
swa_len = tl.minimum(pos + 1, WINDOW_SIZE)
if HAS_IMAGE:
left = tl.load(left_visible_ptr + token_idx)
right = tl.load(right_visible_ptr + token_idx)
else:
left = 0
right = 0
left_add = tl.maximum(left - (WINDOW_SIZE - 1), 0)
# Prefix caching can resume inside an image span. Do not generate
# indices outside the SWA rows present in the gathered workspace.
swa_start = tl.maximum(
tl.maximum(pos - (WINDOW_SIZE - 1) - left_add, 0), gather_start
)
swa_end = tl.minimum(pos + right + 1, seq_len)
swa_len = tl.maximum(swa_end - swa_start, 0)
topk_offset = tl.arange(0, PADDED_TOP_K)
topk_mask = topk_offset < topk_len
@@ -159,14 +182,14 @@ def _combine_topk_swa_indices_kernel(
mask=topk_mask,
)
swa_offset = tl.arange(0, WINDOW_SIZE)
swa_offset = tl.arange(0, PADDED_SWA_WIDTH)
tl.store(
combined_indices_ptr
+ token_idx * combined_indices_stride
+ topk_len
+ swa_offset,
M * batch_idx + N + swa_offset + pos - swa_len + 1 - gather_start,
mask=swa_offset < swa_len,
M * batch_idx + N + swa_offset + swa_start - gather_start,
mask=(swa_offset < swa_len) & (swa_offset < SWA_WIDTH),
)
tl.store(combined_lens_ptr + token_idx, topk_len + swa_len)
@@ -182,12 +205,21 @@ def combine_topk_swa_indices(
topk: int,
M: int,
N: int,
max_image_tokens: int = 0,
left_visible: torch.Tensor | None = None,
right_visible: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor]:
if (left_visible is None) != (right_visible is None):
raise ValueError("left_visible and right_visible must be provided together")
topk_indices = topk_indices.reshape(topk_indices.shape[0], -1).contiguous()
num_tokens = topk_indices.shape[0]
num_reqs = seq_lens.shape[0]
has_image = left_visible is not None
# Keep the row shape fixed for a vision model even when a particular batch
# has no image.
swa_width = window_size + max_image_tokens
combined_topk = (
(topk + window_size + _SPARSE_PREFILL_TOPK_ALIGNMENT - 1)
(topk + swa_width + _SPARSE_PREFILL_TOPK_ALIGNMENT - 1)
// _SPARSE_PREFILL_TOPK_ALIGNMENT
* _SPARSE_PREFILL_TOPK_ALIGNMENT
)
@@ -211,13 +243,18 @@ def combine_topk_swa_indices(
query_start_loc,
seq_lens,
gather_lens,
left_visible if left_visible is not None else topk_indices,
right_visible if right_visible is not None else topk_indices,
M,
N,
TOP_K=topk,
COMPRESS_RATIO=compress_ratio,
WINDOW_SIZE=window_size,
SWA_WIDTH=swa_width,
TOPK_WIDTH=topk_indices.shape[-1],
PADDED_TOP_K=triton.next_power_of_2(topk_indices.shape[-1]),
PADDED_SWA_WIDTH=triton.next_power_of_2(swa_width),
HAS_IMAGE=has_image,
)
return combined_indices, combined_lens
@@ -948,6 +985,12 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
assert query_start_loc_cpu is not None
assert query_start_loc is not None
prefill_token_base = query_start_loc_cpu[num_decodes]
left_visible = swa_metadata.prefill_left_visible
right_visible = swa_metadata.prefill_right_visible
if left_visible is not None:
left_visible = left_visible[num_decode_tokens:]
assert right_visible is not None
right_visible = right_visible[num_decode_tokens:]
if not swa_only:
if self.compress_ratio == 4:
@@ -1026,6 +1069,17 @@ class DeepseekV4ROCMAiterMLAAttention(DeepseekV4Attention):
top_k,
M,
N,
max_image_tokens=self.max_image_tokens,
left_visible=(
left_visible[query_start:query_end]
if left_visible is not None
else None
),
right_visible=(
right_visible[query_start:query_end]
if right_visible is not None
else None
),
)
rocm_sparse_attn_prefill(
q=q[query_start:query_end],
+331
View File
@@ -0,0 +1,331 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""DeepSeek-V4 vision variant (e.g. DeepSeek-V4-Flash-Vision-Exp).
Thin multimodal wrapper around the text-only ``DeepseekV4ForCausalLM``:
- ``vision`` ViT + ``aligner`` produce per-image embeddings for the IMAGE
sentinel positions; four learned vectors (``image_start`` / ``image_pad`` /
``image_newline`` / ``image_end``) fill the remaining sentinel positions.
- Image placeholders (``<|deepseek_image|>``) are expanded by the processor
in ``common/mm_preprocess.py`` into sentinel blocks borrowing reserved
in-vocab tokens ``<|place_holder_mm_span_0431|>``..``_0435|>``
(see ``common/vision.py`` for the tower itself).
- Merged embeddings enter the text model via ``inputs_embeds``, i.e. before
its hyper-connection stream expansion. Raw ``input_ids`` still flow into
the model so the MoE router can apply ``bias_vl`` to image tokens
(``requires_raw_input_tokens``).
"""
from collections.abc import Iterable
import torch
from torch import nn
from vllm.model_executor.models.interfaces import (
MultiModalEmbeddings,
SupportsEagle3,
SupportsMultiModal,
SupportsPP,
)
from vllm.model_executor.models.utils import (
AutoWeightsLoader,
WeightsMapper,
init_vllm_registered_model,
maybe_prefix,
)
from vllm.multimodal import MULTIMODAL_REGISTRY
from .mm_preprocess import (
IMAGE_PLACEHOLDER,
IMAGE_SENTINEL_BASE_ID,
DeepseekV4VLDummyInputsBuilder,
DeepseekV4VLMultiModalProcessor,
DeepseekV4VLProcessingInfo,
image_sentinel_mask,
)
from .vision import DeepseekV4Aligner, DeepseekV4ViT
def _make_deepseek_v4_vl_weights_mapper(
text_mapper: WeightsMapper, image_enabled: bool
) -> WeightsMapper:
"""Text-checkpoint mapping rules re-rooted under ``language_model.``."""
orig_to_new_prefix = {
src: None if dst is None else f"language_model.{dst}"
for src, dst in text_mapper.orig_to_new_prefix.items()
}
if not image_enabled:
orig_to_new_prefix.update({"vision.": None, "aligner.": None, "image_": None})
return WeightsMapper(
orig_to_new_renaming=text_mapper.orig_to_new_renaming,
orig_to_new_prefix=orig_to_new_prefix,
orig_to_new_regex=text_mapper.orig_to_new_regex,
orig_to_new_stacked=text_mapper.orig_to_new_stacked,
orig_to_new_suffix={
**text_mapper.orig_to_new_suffix,
"head.weight": "language_model.lm_head.weight",
},
orig_to_new_substr={
**text_mapper.orig_to_new_substr,
# Draft models load the checkpoint's MTP weights separately from
# the target model.
"mtp.": None,
},
)
@MULTIMODAL_REGISTRY.register_processor(
DeepseekV4VLMultiModalProcessor,
info=DeepseekV4VLProcessingInfo,
dummy_inputs=DeepseekV4VLDummyInputsBuilder,
)
class DeepseekV4ForConditionalGeneration(
nn.Module, SupportsMultiModal, SupportsPP, SupportsEagle3
):
"""Multimodal entry point for DeepSeek-V4 checkpoints with a vision tower.
``SupportsEagle3`` (aux hidden-state plumbing for MTP/DSpark drafters)
delegates through ``language_model`` via the protocol defaults.
"""
# The MoE router needs raw token ids to detect image sentinel tokens
# (borrowed reserved ids, see common/mm_preprocess.py) and apply bias_vl.
requires_raw_input_tokens = True
@classmethod
def get_placeholder_str(cls, modality: str, i: int) -> str | None:
if modality == "image":
return IMAGE_PLACEHOLDER
raise ValueError(f"Unsupported modality: {modality!r}")
def __init__(self, *, vllm_config, prefix: str = "") -> None:
super().__init__()
model_config = vllm_config.model_config
config = model_config.hf_config
self.config = config
self.multimodal_config = model_config.multimodal_config
assert self.multimodal_config is not None
image_enabled = (
config.vision_n_layers > 0
and self.multimodal_config.get_limit_per_prompt("image") > 0
)
with self._mark_tower_model(vllm_config, {"image"}):
self.vision: DeepseekV4ViT | None = None
self.aligner: DeepseekV4Aligner | None = None
self.image_start: nn.Parameter | None = None
self.image_end: nn.Parameter | None = None
self.image_newline: nn.Parameter | None = None
self.image_pad: nn.Parameter | None = None
if image_enabled:
self.vision = DeepseekV4ViT(config)
self.aligner = DeepseekV4Aligner(config)
for name in (
"image_start",
"image_end",
"image_newline",
"image_pad",
):
setattr(
self,
name,
nn.Parameter(
torch.empty(config.hidden_size, dtype=torch.float32)
),
)
self.vision.to(dtype=model_config.dtype)
self.aligner.to(dtype=model_config.dtype)
with self._mark_language_model(vllm_config):
# The arch convertor routes any config with a vision tower to
# this wrapper class; mark the copy handed to the text backbone
# so it resolves to DeepseekV4ForCausalLM instead of recursing
# (with_hf_config deepcopies the config, the marker survives).
config._dsv4_vl_inner = True # type: ignore[attr-defined]
try:
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["DeepseekV4ForCausalLM"],
)
finally:
del config._dsv4_vl_inner # type: ignore[attr-defined]
# The outer mapper (see load_weights) fully resolves HF names into
# this wrapper's namespace before AutoWeightsLoader strips the
# "language_model." prefix and delegates to the child's load_weights,
# so the child's own mapper must be a no-op. Its suffix rules are not
# idempotent (e.g. "lm_head.weight".endswith("head.weight") would
# re-fire "head.weight" -> "lm_head.weight").
text_mapper = self.language_model.hf_to_vllm_mapper
self.language_model.hf_to_vllm_mapper = WeightsMapper()
self.make_empty_intermediate_tensors = ( # type: ignore[method-assign]
self.language_model.make_empty_intermediate_tensors
)
self.hf_to_vllm_mapper = _make_deepseek_v4_vl_weights_mapper(
text_mapper, image_enabled
)
self._weights_finalized = False
def _parse_and_validate_image_input(self, **kwargs: object) -> dict | None:
patches = kwargs.pop("patches", None)
if patches is None:
return None
vit_grid = kwargs.pop("vit_grid", None)
llm_grid = kwargs.pop("llm_grid", None)
perm = kwargs.pop("perm", None)
assert vit_grid is not None and llm_grid is not None and perm is not None
return {
"patches": patches,
"vit_grid": vit_grid,
"llm_grid": llm_grid,
"perm": perm,
}
def _encode_image(
self,
patches: torch.Tensor,
n_vit_h: int,
n_vit_w: int,
perm: torch.Tensor,
) -> torch.Tensor:
assert self.vision is not None and self.aligner is not None
image_embeds = self.aligner(
self.vision(patches, n_vit_h, n_vit_w), n_vit_h, n_vit_w
)
# Reorder into the N-layout block order used in the prompt.
return image_embeds[perm.to(image_embeds.device)]
def _process_image_input(
self,
patches: torch.Tensor,
vit_grid: torch.Tensor,
llm_grid: torch.Tensor,
perm: torch.Tensor,
) -> tuple[torch.Tensor, ...]:
assert self.vision is not None and self.aligner is not None
patches = patches.to(self.aligner.w1.weight.dtype)
embeds: list[torch.Tensor] = []
vit_offset = 0
llm_offset = 0
for (n_vit_h, n_vit_w), (n_llm_h, n_llm_w) in zip(
vit_grid.tolist(), llm_grid.tolist(), strict=True
):
n_vit = n_vit_h * n_vit_w
n_llm = n_llm_h * n_llm_w
embeds.append(
self._encode_image(
patches[vit_offset : vit_offset + n_vit],
n_vit_h,
n_vit_w,
perm[llm_offset : llm_offset + n_llm],
)
)
vit_offset += n_vit
llm_offset += n_llm
return tuple(embeds)
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
image_input = self._parse_and_validate_image_input(**kwargs)
if image_input is None or self.vision is None:
return []
return self._process_image_input(
image_input["patches"],
image_input["vit_grid"],
image_input["llm_grid"],
image_input["perm"],
)
def embed_input_ids(
self,
input_ids: torch.Tensor,
multimodal_embeddings: MultiModalEmbeddings | None = None,
*,
is_multimodal: torch.Tensor | None = None,
) -> torch.Tensor:
from vllm.model_executor.models.utils import _merge_multimodal_embeddings
# All ids are in-vocab here: image-block sentinels are borrowed
# reserved tokens (their embedding rows are always overwritten below).
inputs_embeds = self.language_model.embed_input_ids(input_ids)
if self.image_start is not None:
# Branch-free sentinel overwrite: safe inside compiled/captured
# regions (no data-dependent control flow).
sentinel_mask = image_sentinel_mask(input_ids)
if is_multimodal is not None:
# IMAGE positions get vision embeddings via the merge below.
sentinel_mask = sentinel_mask & ~is_multimodal.to(input_ids.device)
table = torch.stack(
[
self.image_start,
self.image_pad,
self.image_pad,
self.image_newline,
self.image_end,
]
).to(inputs_embeds.dtype)
idx = (input_ids - IMAGE_SENTINEL_BASE_ID).clamp(0, 4)
inputs_embeds = torch.where(
sentinel_mask.unsqueeze(-1), table[idx], inputs_embeds
)
if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
return inputs_embeds
assert is_multimodal is not None
return _merge_multimodal_embeddings(
inputs_embeds=inputs_embeds,
multimodal_embeddings=multimodal_embeddings,
is_multimodal=is_multimodal,
)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors=None,
inputs_embeds: torch.Tensor | None = None,
**kwargs,
) -> torch.Tensor:
return self.language_model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
return self.language_model.compute_logits(hidden_states)
def compute_logits_local(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.language_model.compute_logits_local(hidden_states)
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return self.language_model.get_expert_mapping()
def get_mtp_target_hidden_states(self) -> torch.Tensor | None:
"""Pre-hc_head residual stream buffer for the MTP/DSpark draft model."""
return self.language_model.get_mtp_target_hidden_states()
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
child_finalizes = getattr(
self.language_model, "finalizes_weights_during_load", True
)
mapped = self.hf_to_vllm_mapper.apply(weights)
if child_finalizes:
# A child which finalizes inside load_weights must see all of its
# weights in one contiguous delegation from AutoWeightsLoader.
mapped = iter(sorted(mapped, key=lambda x: x[0]))
loader = AutoWeightsLoader(self)
loaded_params = loader.load_weights(mapped)
self._weights_finalized = child_finalizes
return loaded_params
def process_weights_after_loading(self) -> None:
# Backbones such as the ROCm implementation require this to run only
# after the loader's generic per-layer quantization finalization.
if getattr(self, "_weights_finalized", False):
return
self.language_model.process_weights_after_loading()
self._weights_finalized = True
+1
View File
@@ -1790,6 +1790,7 @@ class DeepseekV4ForCausalLM(
SupportsLoRA,
DeepseekV4MixtureOfExperts,
):
finalizes_weights_during_load = True
model_cls = DeepseekV4Model
# Default mapper assumes the original FP4-expert checkpoint layout.
+10 -322
View File
@@ -1,333 +1,21 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""DeepSeek-V4 vision variant (e.g. DeepSeek-V4-Flash-Vision-Exp).
"""Compatibility imports for the platform-neutral DeepSeek-V4 vision model."""
Thin multimodal wrapper around the text-only ``DeepseekV4ForCausalLM``:
- ``vision`` ViT + ``aligner`` produce per-image embeddings for the IMAGE
sentinel positions; four learned vectors (``image_start`` / ``image_pad`` /
``image_newline`` / ``image_end``) fill the remaining sentinel positions.
- Image placeholders (``<|deepseek_image|>``) are expanded by the processor
in ``common/mm_preprocess.py`` into sentinel blocks borrowing reserved
in-vocab tokens ``<|place_holder_mm_span_0431|>``..``_0435|>``
(see ``common/vision.py`` for the tower itself).
- Merged embeddings enter the text model via ``inputs_embeds``, i.e. before
its hyper-connection stream expansion. Raw ``input_ids`` still flow into
the model so the MoE router can apply ``bias_vl`` to image tokens
(``requires_raw_input_tokens``).
"""
from collections.abc import Iterable
import torch
from torch import nn
from vllm.model_executor.models.interfaces import (
MultiModalEmbeddings,
SupportsEagle3,
SupportsMultiModal,
SupportsPP,
from ..common.vl_model import (
DeepseekV4ForConditionalGeneration,
)
from vllm.model_executor.models.utils import (
AutoWeightsLoader,
WeightsMapper,
init_vllm_registered_model,
maybe_prefix,
from ..common.vl_model import (
_make_deepseek_v4_vl_weights_mapper as _make_common_vl_weights_mapper,
)
from vllm.multimodal import MULTIMODAL_REGISTRY
from ..common.mm_preprocess import (
IMAGE_PLACEHOLDER,
IMAGE_SENTINEL_BASE_ID,
DeepseekV4VLDummyInputsBuilder,
DeepseekV4VLMultiModalProcessor,
DeepseekV4VLProcessingInfo,
image_sentinel_mask,
)
from ..common.vision import DeepseekV4Aligner, DeepseekV4ViT
from .model import _make_deepseek_v4_weights_mapper
def _make_deepseek_v4_vl_weights_mapper(
expert_dtype: str, image_enabled: bool
) -> WeightsMapper:
"""Text-checkpoint mapping rules re-rooted under ``language_model.``."""
base = _make_deepseek_v4_weights_mapper(expert_dtype)
orig_to_new_prefix: dict[str, str | None] = {
"layers.": "language_model.model.layers.",
"embed.": "language_model.model.embed.",
"norm.": "language_model.model.norm.",
"hc_head": "language_model.model.hc_head",
"mtp.": "language_model.model.mtp.",
}
if not image_enabled:
orig_to_new_prefix.update({"vision.": None, "aligner.": None, "image_": None})
return WeightsMapper(
orig_to_new_prefix=orig_to_new_prefix,
orig_to_new_regex=base.orig_to_new_regex,
orig_to_new_suffix={
"head.weight": "language_model.lm_head.weight",
"embed.weight": "embed_tokens.weight",
".ffn.gate.bias": ".ffn.gate.e_score_correction_bias",
},
orig_to_new_substr={
".shared_experts.w2": ".shared_experts.down_proj",
# The MTP/DSpark draft heads are not supported for the vision
# variant; drop their weights.
"mtp.": None,
},
def _make_deepseek_v4_vl_weights_mapper(expert_dtype: str, image_enabled: bool):
"""Retain the original NVIDIA helper signature for downstream imports."""
return _make_common_vl_weights_mapper(
_make_deepseek_v4_weights_mapper(expert_dtype), image_enabled
)
@MULTIMODAL_REGISTRY.register_processor(
DeepseekV4VLMultiModalProcessor,
info=DeepseekV4VLProcessingInfo,
dummy_inputs=DeepseekV4VLDummyInputsBuilder,
)
class DeepseekV4ForConditionalGeneration(
nn.Module, SupportsMultiModal, SupportsPP, SupportsEagle3
):
"""Multimodal entry point for DeepSeek-V4 checkpoints with a vision tower.
``SupportsEagle3`` (aux hidden-state plumbing for MTP/DSpark drafters)
delegates through ``language_model`` via the protocol defaults.
"""
# The MoE router needs raw token ids to detect image sentinel tokens
# (borrowed reserved ids, see common/mm_preprocess.py) and apply bias_vl.
requires_raw_input_tokens = True
@classmethod
def get_placeholder_str(cls, modality: str, i: int) -> str | None:
if modality == "image":
return IMAGE_PLACEHOLDER
raise ValueError(f"Unsupported modality: {modality!r}")
def __init__(self, *, vllm_config, prefix: str = "") -> None:
super().__init__()
model_config = vllm_config.model_config
config = model_config.hf_config
self.config = config
self.multimodal_config = model_config.multimodal_config
assert self.multimodal_config is not None
image_enabled = (
config.vision_n_layers > 0
and self.multimodal_config.get_limit_per_prompt("image") > 0
)
with self._mark_tower_model(vllm_config, {"image"}):
self.vision: DeepseekV4ViT | None = None
self.aligner: DeepseekV4Aligner | None = None
self.image_start: nn.Parameter | None = None
self.image_end: nn.Parameter | None = None
self.image_newline: nn.Parameter | None = None
self.image_pad: nn.Parameter | None = None
if image_enabled:
self.vision = DeepseekV4ViT(config)
self.aligner = DeepseekV4Aligner(config)
for name in (
"image_start",
"image_end",
"image_newline",
"image_pad",
):
setattr(
self,
name,
nn.Parameter(
torch.empty(config.hidden_size, dtype=torch.float32)
),
)
self.vision.to(dtype=model_config.dtype)
self.aligner.to(dtype=model_config.dtype)
with self._mark_language_model(vllm_config):
# The arch convertor routes any config with a vision tower to
# this wrapper class; mark the copy handed to the text backbone
# so it resolves to DeepseekV4ForCausalLM instead of recursing
# (with_hf_config deepcopies the config, the marker survives).
config._dsv4_vl_inner = True # type: ignore[attr-defined]
try:
self.language_model = init_vllm_registered_model(
vllm_config=vllm_config,
hf_config=config,
prefix=maybe_prefix(prefix, "language_model"),
architectures=["DeepseekV4ForCausalLM"],
)
finally:
del config._dsv4_vl_inner # type: ignore[attr-defined]
# The outer mapper (see load_weights) fully resolves HF names into
# this wrapper's namespace before AutoWeightsLoader strips the
# "language_model." prefix and delegates to the child's load_weights,
# so the child's own mapper must be a no-op. Its suffix rules are not
# idempotent (e.g. "lm_head.weight".endswith("head.weight") would
# re-fire "head.weight" -> "lm_head.weight").
self.language_model.hf_to_vllm_mapper = WeightsMapper()
self.make_empty_intermediate_tensors = ( # type: ignore[method-assign]
self.language_model.make_empty_intermediate_tensors
)
expert_dtype = getattr(config, "expert_dtype", "fp4")
self.hf_to_vllm_mapper = _make_deepseek_v4_vl_weights_mapper(
expert_dtype, image_enabled
)
def _parse_and_validate_image_input(self, **kwargs: object) -> dict | None:
patches = kwargs.pop("patches", None)
if patches is None:
return None
vit_grid = kwargs.pop("vit_grid", None)
llm_grid = kwargs.pop("llm_grid", None)
perm = kwargs.pop("perm", None)
assert vit_grid is not None and llm_grid is not None and perm is not None
return {
"patches": patches,
"vit_grid": vit_grid,
"llm_grid": llm_grid,
"perm": perm,
}
def _encode_image(
self,
patches: torch.Tensor,
n_vit_h: int,
n_vit_w: int,
perm: torch.Tensor,
) -> torch.Tensor:
assert self.vision is not None and self.aligner is not None
image_embeds = self.aligner(
self.vision(patches, n_vit_h, n_vit_w), n_vit_h, n_vit_w
)
# Reorder into the N-layout block order used in the prompt.
return image_embeds[perm.to(image_embeds.device)]
def _process_image_input(
self,
patches: torch.Tensor,
vit_grid: torch.Tensor,
llm_grid: torch.Tensor,
perm: torch.Tensor,
) -> tuple[torch.Tensor, ...]:
assert self.vision is not None and self.aligner is not None
patches = patches.to(self.aligner.w1.weight.dtype)
embeds: list[torch.Tensor] = []
vit_offset = 0
llm_offset = 0
for (n_vit_h, n_vit_w), (n_llm_h, n_llm_w) in zip(
vit_grid.tolist(), llm_grid.tolist(), strict=True
):
n_vit = n_vit_h * n_vit_w
n_llm = n_llm_h * n_llm_w
embeds.append(
self._encode_image(
patches[vit_offset : vit_offset + n_vit],
n_vit_h,
n_vit_w,
perm[llm_offset : llm_offset + n_llm],
)
)
vit_offset += n_vit
llm_offset += n_llm
return tuple(embeds)
def embed_multimodal(self, **kwargs: object) -> MultiModalEmbeddings:
image_input = self._parse_and_validate_image_input(**kwargs)
if image_input is None or self.vision is None:
return []
return self._process_image_input(
image_input["patches"],
image_input["vit_grid"],
image_input["llm_grid"],
image_input["perm"],
)
def embed_input_ids(
self,
input_ids: torch.Tensor,
multimodal_embeddings: MultiModalEmbeddings | None = None,
*,
is_multimodal: torch.Tensor | None = None,
) -> torch.Tensor:
from vllm.model_executor.models.utils import _merge_multimodal_embeddings
# All ids are in-vocab here: image-block sentinels are borrowed
# reserved tokens (their embedding rows are always overwritten below).
inputs_embeds = self.language_model.embed_input_ids(input_ids)
if self.image_start is not None:
# Branch-free sentinel overwrite: safe inside compiled/captured
# regions (no data-dependent control flow).
sentinel_mask = image_sentinel_mask(input_ids)
if is_multimodal is not None:
# IMAGE positions get vision embeddings via the merge below.
sentinel_mask = sentinel_mask & ~is_multimodal.to(input_ids.device)
table = torch.stack(
[
self.image_start,
self.image_pad,
self.image_pad,
self.image_newline,
self.image_end,
]
).to(inputs_embeds.dtype)
idx = (input_ids - IMAGE_SENTINEL_BASE_ID).clamp(0, 4)
inputs_embeds = torch.where(
sentinel_mask.unsqueeze(-1), table[idx], inputs_embeds
)
if multimodal_embeddings is None or len(multimodal_embeddings) == 0:
return inputs_embeds
assert is_multimodal is not None
return _merge_multimodal_embeddings(
inputs_embeds=inputs_embeds,
multimodal_embeddings=multimodal_embeddings,
is_multimodal=is_multimodal,
)
def forward(
self,
input_ids: torch.Tensor,
positions: torch.Tensor,
intermediate_tensors=None,
inputs_embeds: torch.Tensor | None = None,
**kwargs,
) -> torch.Tensor:
return self.language_model(
input_ids, positions, intermediate_tensors, inputs_embeds
)
def compute_logits(self, hidden_states: torch.Tensor) -> torch.Tensor | None:
return self.language_model.compute_logits(hidden_states)
def compute_logits_local(self, hidden_states: torch.Tensor) -> torch.Tensor:
return self.language_model.compute_logits_local(hidden_states)
def get_expert_mapping(self) -> list[tuple[str, str, int, str]]:
return self.language_model.get_expert_mapping()
def get_mtp_target_hidden_states(self) -> torch.Tensor | None:
"""Pre-hc_head residual stream buffer for the MTP/DSpark draft model."""
return self.language_model.get_mtp_target_hidden_states()
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]) -> set[str]:
# Map HF names into this wrapper's namespace up front and sort, so
# the "language_model." group reaches the child loader as one
# contiguous block (AutoWeightsLoader delegates per contiguous group,
# and the child's load_weights finalizes fused expert weights, which
# must not run on a partially loaded model).
mapped = sorted(self.hf_to_vllm_mapper.apply(weights), key=lambda x: x[0])
loader = AutoWeightsLoader(self)
loaded_params = loader.load_weights(mapped)
# The child's load_weights already ran its post-load finalization.
self._weights_finalized = True
return loaded_params
def process_weights_after_loading(self) -> None:
# Model-level post-load hook (called by the loader after any load
# format). Under DummyModelLoader the child's load_weights — and
# hence its finalize step — is bypassed, so run it here instead.
if getattr(self, "_weights_finalized", False):
return
self.language_model.process_weights_after_loading()
__all__ = ["DeepseekV4ForConditionalGeneration"]
+2 -2
View File
@@ -9,6 +9,6 @@ class DeepseekV4ForConditionalGeneration(nn.Module):
def __init__(self, *, vllm_config, prefix: str = ""):
super().__init__()
raise NotImplementedError(
"DeepSeek-V4 vision (DeepseekV4ForConditionalGeneration) is only "
"supported on NVIDIA GPUs for now."
"DeepSeek-V4 vision (DeepseekV4ForConditionalGeneration) is not "
"supported on the current platform; use CUDA or ROCm."
)
+15
View File
@@ -943,6 +943,21 @@ class RocmPlatform(Platform):
if parallel_config.worker_cls == "auto":
parallel_config.worker_cls = "vllm.v1.worker.gpu_worker.Worker"
model_config = vllm_config.model_config
scheduler_config = vllm_config.scheduler_config
# Note: model_config may be None during testing
if (
model_config is not None
and model_config.is_mm_prefix_lm
and scheduler_config.is_multimodal_model
and not scheduler_config.disable_chunked_mm_input
):
logger.warning_once(
"Forcing --disable_chunked_mm_input for models "
"with multimodal-bidirectional attention."
)
scheduler_config.disable_chunked_mm_input = True
@classmethod
def verify_model_arch(cls, model_arch: str) -> None:
if model_arch in _ROCM_UNSUPPORTED_MODELS: