[Bugfix] Add seed support to TTS API for deterministic Fish Speech voice generation (#2624)

Signed-off-by: Yiyang Liu <yiyangliu@microsoft.com>
Signed-off-by: Yiyang Liu <37043548+ianliuy@users.noreply.github.com>
Signed-off-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Yueqian Lin <70319226+linyueqian@users.noreply.github.com>
This commit is contained in:
authored and GitHub committed 2026-04-26 19:15:30 +00:00
1 parent 8e2e7c74fb
commit 62c86c9d62
7 files changed
+94 -9

No files matched your search

+30 -4
View File
@@ -1372,6 +1372,35 @@ class OmniGPUModelRunner(GPUModelRunner):
subtalker_params = getattr(self.vllm_config.model_config, "subtalker_sampling_params", None)
if not isinstance(subtalker_params, dict):
subtalker_params = {}
# Extract seed from the first request's sampling params for Fast AR
# determinism. NOTE: when batch_size > 1, all requests share the first
# request's seed. Per-request Fast AR seeding requires row-by-row
# torch.multinomial calls and is left as a follow-up optimisation.
_seed = None
if decode_req_ids:
_first_sp = getattr(self.requests[decode_req_ids[0]], "sampling_params", None)
if _first_sp is not None and getattr(_first_sp, "seed", None) is not None:
_seed = _first_sp.seed
# Warn when batched requests have different seeds.
if len(decode_req_ids) > 1 and _seed is not None:
_other_seeds = {
getattr(getattr(self.requests[rid], "sampling_params", None), "seed", None)
for rid in decode_req_ids[1:]
}
if _other_seeds != {_seed}:
logger.warning(
"Fast AR seed: batch has mixed seeds; using first request's seed=%d for all %d requests.",
_seed,
len(decode_req_ids),
)
talker_kwargs = {
"do_sample": subtalker_params.get("do_sample"),
"temperature": subtalker_params.get("temperature"),
"top_k": subtalker_params.get("top_k"),
"top_p": subtalker_params.get("top_p"),
}
if _seed is not None:
talker_kwargs["seed"] = _seed
with set_forward_context(
None, self.vllm_config, cudagraph_runtime_mode=_cudagraph_mode, batch_descriptor=batch_desc
):
@@ -1380,10 +1409,7 @@ class OmniGPUModelRunner(GPUModelRunner):
req_embeds,
last_talker_hidden,
text_step,
do_sample=subtalker_params.get("do_sample"),
temperature=subtalker_params.get("temperature"),
top_k=subtalker_params.get("top_k"),
top_p=subtalker_params.get("top_p"),
**talker_kwargs,
)
# update the inputs_embeds and code_predictor_codes
out_key = getattr(self.model, "talker_mtp_output_key", ("codes", "audio"))