[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
@@ -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