Correct GLM-TTS model_specs_v1 metadata

[mirek190]

Summary

Follow-up to #98 that aligns GLM-TTS with the normalized model_specs_v1
option contract while preserving its existing request keys as compatibility
aliases.

Changes

• keeps the normalized v1 request names:
  num_inference_steps, flow_guidance_scale, flow_noise_path,
  hift_source_random_path, and hift_prior_noise_count
• advertises those normalized names through the loader
• maps normalized names to the existing GLM-TTS implementation
• retains flow_steps, cfg_rate, flow_noise_file,
  hift_source_random_file, and hift_prior_noise_values as backward-compatible
  aliases
• keeps language metadata scoped to Chinese and English
• removes claims for streaming and explicit phoneme/emotion controls that the
  current native runtime does not expose
• makes Llama context-size minimums match runtime validation
• describes the standalone model accurately as mixed Q8_0/F16
• installs the package into the documented and tested GLM-TTS-Q8 directory
• preserves the existing SafeTensors/GGUF source layouts

Why

model_specs_v1 is the normalized public contract used by framework validation,
catalog clients, and generated UIs. It must not mirror model-specific legacy
implementation names. Compatibility belongs at the model boundary, so new
clients get one consistent vocabulary without breaking existing GLM-TTS
commands.

Impact

Normalized v1 requests now work directly. Existing requests using the old
GLM-TTS keys continue to work. The option resolver also rejects conflicting
values when both a normalized name and its legacy alias are supplied.

Inference math, weights, package loading, and generated output are unchanged.

Validation

• parsed model_specs_v1/glm_tts.json
• confirmed all normalized request names against the framework contracts
• confirmed v1 sources remain identical to model_specs/glm_tts.json
• resolved glm_tts_q8_0 with model_manager_v2.py
• verified the remote GGUF at 5,143,813,728 bytes
• built the complete engine_runtime CPU target
• built glm_tts_warm_bench with normalized option ingestion
• git diff --check

[mirek190]

This is a metadata-only follow-up to #98, based on current main 41246aa.

Why this update was needed

The migrated GLM-TTS v1 entry contained five descriptive option aliases that are not the public keys consumed by GlmTTSSession. A catalog/UI generated from that entry would therefore send ignored options.

Previous v1 name	Runtime key
num_inference_steps	flow_steps
flow_guidance_scale	cfg_rate
flow_noise_path	flow_noise_file
hift_source_random_path	hift_source_random_file
hift_prior_noise_count	hift_prior_noise_values

The description also mixed upstream GLM-TTS capabilities with what audio.cpp currently exposes. Upstream supports streaming and optional phoneme input, but the native audio.cpp contribution currently advertises offline tts/clone and accepts ordinary text plus reference audio/transcript. The spec now describes that implemented surface only. Chinese and English remain the correct language metadata.

Package metadata corrections

The published standalone file is the quality-tested mixed model: the autoregressive Llama tensors remain F16 while speech-tokenizer, Flow, HiFT, and CAMPPlus tensors are Q8_0. The display name and package description now say this explicitly. Its install target is corrected from GLM-TTS-GGUF to GLM-TTS-Q8, matching the existing documentation and validation paths.

The SafeTensors and GGUF sources arrays are byte-for-byte structurally identical to model_specs/glm_tts.json, and no schema_version field was added.

Validation

• JSON parse: pass
• v1/runtime-bridge source-layout equality: pass
• tasks: tts, clone
• languages: zh, en
• model_manager_v2.py info glm_tts: pass
• remote package check: pass (5,143,813,728 bytes)
• git diff --check: pass

No C++ inference code changed, so audio generation behavior and existing GGUF compatibility are unaffected.

[0xShug0]

@mirek190 Sorry, I didn't explain this clearly. The v1 spec uses normalized option keys so that different models don’t use different names for the same concept. It also allows the framework to validate certain options centrally (topk etc), instead of duplicating validation logic in every model.

When we migrate existing models, we can add a compatibility layer for each model to map the old keys in the requests to the new normalized keys, or just directly align model names with normalized v1 keys. New models should use the normalized keys directly, or at least follow the same naming conventions for any model-specific option keys.

I’m currently working on releasing a model early to serve as a reference implementation for future v1-native new models.

[mirek190]

Thanks for the clarification. Commit 81a2c28 revises this PR to follow the intended v1 direction.

• Restored the normalized spec names: num_inference_steps, flow_guidance_scale, flow_noise_path, hift_source_random_path, and hift_prior_noise_count.
• Updated loader metadata and GLM-TTS documentation to advertise those normalized names.
• Added an additive model-boundary compatibility mapping so the existing flow_steps, cfg_rate, flow_noise_file, hift_source_random_file, and hift_prior_noise_values keys continue to work.
• Normalized keys are checked first; if both forms are supplied with different values, the shared resolver reports a conflict rather than choosing one silently.
• Updated the GLM warm-bench input path to exercise the normalized names.

Validation: JSON/model-manager checks pass, v1/runtime source layouts remain equal, the complete CPU engine_runtime target builds, glm_tts_warm_bench builds, and git diff --check passes. Inference math and generated outputs are unchanged.