[2/N][KV-Cache Layout Refactor] Pack K/V into the content dim across attention backends (#44455)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
Signed-off-by: Lucas Wilkinson <LucasWilkinson@users.noreply.github.com>
Co-authored-by: Claude <noreply@anthropic.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Matthew Bonanni <mbonanni@redhat.com>
Co-authored-by: Nicolò Lucchesi <nlucches@redhat.com>
This commit is contained in:
Lucas Wilkinson
2026-07-11 11:11:16 -04:00
committed by GitHub
co-authored by Claude OpenAI Codex Matthew Bonanni Nicolò Lucchesi
parent 76fedaa2a5
commit 51878e5b6e
31 changed files with 664 additions and 604 deletions
+3 -5
View File
@@ -10,7 +10,7 @@ from tests.kernels.utils import DEFAULT_OPCHECK_TEST_UTILS, opcheck
from vllm import _custom_ops as ops
from vllm.model_executor.layers.quantization.utils.quant_utils import scaled_dequantize
from vllm.platforms import current_platform
from vllm.utils.torch_utils import nvfp4_kv_cache_split_views, set_random_seed
from vllm.utils.torch_utils import nvfp4_split_data_scale, set_random_seed
COPYING_DIRECTION = [("cuda", "cpu"), ("cuda", "cuda"), ("cpu", "cuda")]
DTYPES = [torch.bfloat16, torch.float]
@@ -255,10 +255,8 @@ def test_reshape_and_cache_flash(
nvfp4_key_data = None
nvfp4_value_data = None
if kv_cache_dtype == "nvfp4":
(nvfp4_key_data,), (key_scale_cache,) = nvfp4_kv_cache_split_views(key_cache)
(nvfp4_value_data,), (value_scale_cache,) = nvfp4_kv_cache_split_views(
value_cache
)
nvfp4_key_data, key_scale_cache = nvfp4_split_data_scale(key_cache)
nvfp4_value_data, value_scale_cache = nvfp4_split_data_scale(value_cache)
if kv_cache_dtype == "nvfp4":
# Global scale = amax / 448 (per-tensor)