Skip to content

[CT] Fix CT Config to honor fp8_inc KV cache dtype #929

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
yiliu30 wants to merge 3 commits into
base: main
Choose a base branch
from
+38 −2
Open

[CT] Fix CT Config to honor fp8_inc KV cache dtype #929

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
60223ea
Revert LLMC override
yiliu30 Feb 4, 2026
329c363
fix
yiliu30 Feb 4, 2026
6b0eac3
Merge branch 'main' into fix-llmc-kv
yiliu30 Feb 5, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
40 changes: 38 additions & 2 deletions vllm_gaudi/ops/hpu_compressed_tensors.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -14,8 +14,13 @@
PackedvLLMParameter, RowvLLMParameter)
from vllm.model_executor.layers.quantization.compressed_tensors import (compressed_tensors)
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import ( # noqa: E501
CompressedTensorsLinearMethod as OrigCompressedTensorsLinearMethod, CompressedTensorsConfig,
CompressedTensorsMoEMethod, CompressedTensorsKVCacheMethod)
CompressedTensorsLinearMethod as OrigCompressedTensorsLinearMethod)
from vllm.model_executor.layers.quantization.compressed_tensors.compressed_tensors import (
CompressedTensorsConfig,
CompressedTensorsMoEMethod,
CompressedTensorsKVCacheMethod,
SparsityCompressionConfig,
)
from vllm.model_executor.layers.quantization.compressed_tensors import (compressed_tensors_moe)
from vllm.model_executor.layers.quantization.compressed_tensors.schemes import ( # noqa: E501
CompressedTensorsScheme, CompressedTensorsWNA16)
Expand Down Expand Up @@ -807,6 +812,37 @@ def process_weights_after_loading(self, layer: torch.nn.Module) -> None:

class HPUCompressedTensorsConfig(CompressedTensorsConfig):

def __init__(
self,
target_scheme_map: dict[str, Any],
ignore: list[str],
quant_format: str,
sparsity_scheme_map: dict[str, SparsityCompressionConfig],
sparsity_ignore_list: list[str],
kv_cache_scheme: dict[str, Any] | None = None,
config: dict[str, Any] | None = None,
transform_config: dict[str, Any] | None = None,
total_num_heads: int | None = None,
total_num_kv_heads: int | None = None,
):
super().__init__(
target_scheme_map,
ignore,
quant_format,
sparsity_scheme_map,
sparsity_ignore_list,
kv_cache_scheme,
config,
transform_config,
total_num_heads,
total_num_kv_heads,
)
# Fix https://github.com/vllm-project/vllm/pull/30141
# LLMC overrides the `kv_cache_dtype` to 'fp8', while HPU uses 'fp8_inc'.
if getattr(self, "kv_cache_scheme", None) is not None:
self.kv_cache_dtype = "fp8_inc"
self.kv_cache_scheme = None

def get_quant_method(
self,
layer: torch.nn.Module,
Expand Down