Skip to content

feat: add KV cache quantization args to server #1073

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
deceptech-packet-ninja wants to merge 2 commits into
base: main
Choose a base branch
from
Open

feat: add KV cache quantization args to server #1073

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
5778fdb
feat: add --kv-bits, --kv-group-size, --quantized-kv-start to server
deceptech-packet-ninja Mar 30, 2026
5a67289
merge: resolve conflict with upstream batch generation refactor
deceptech-packet-ninja Apr 3, 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
26 changes: 26 additions & 0 deletions mlx_lm/server.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -671,6 +671,8 @@ def _make_state_machine(
return sm, sequences

def _is_batchable(self, args):
if getattr(self.cli_args, "kv_bits", None) is not None:
return False
return self.model_provider.is_batchable and args.seed is None

def _generate(self):
Expand Down Expand Up @@ -953,6 +955,11 @@ def progress(tokens_processed, tokens_total):
cache += make_prompt_cache(self.model_provider.draft_model)

# Process the prompt and generate tokens
kv_kwargs = {}
if getattr(self.cli_args, "kv_bits", None) is not None:
kv_kwargs["kv_bits"] = self.cli_args.kv_bits
kv_kwargs["kv_group_size"] = self.cli_args.kv_group_size
kv_kwargs["quantized_kv_start"] = self.cli_args.quantized_kv_start
for gen in stream_generate(
model=model,
tokenizer=tokenizer,
Expand All @@ -965,6 +972,7 @@ def progress(tokens_processed, tokens_total):
num_draft_tokens=args.num_draft_tokens,
prompt_progress_callback=progress,
prefill_step_size=self.cli_args.prefill_step_size,
**kv_kwargs,
):
finish_reason = gen.finish_reason
sm_state, match_sequence, current_state = sm.match(sm_state, gen.token)
Expand Down Expand Up @@ -1861,6 +1869,24 @@ def main():
default=2048,
help="Step size for prefill processing (default: 2048)",
)
parser.add_argument(
"--kv-bits",
type=int,
default=None,
help="Number of bits for KV cache quantization. None means no quantization.",
)
parser.add_argument(
"--kv-group-size",
type=int,
default=64,
help="Group size for KV cache quantization (default: 64)",
)
parser.add_argument(
"--quantized-kv-start",
type=int,
default=0,
help="Step to begin quantizing the KV cache (default: 0)",
)
parser.add_argument(
"--prompt-cache-size",
type=int,
Expand Down