server: do not propagate --kv-mean-center bias to draft model contexts

[cnndabbler]

Problem

--kv-mean-center sets common_params.kv_mean_center_path globally. The server builds draft-model params by copying the whole struct (auto params_dft = params_base;), so when a draft model is loaded via -md, both the draft VRAM-measurement probe and the real draft context inherit the bias path. The bias GGUF is calibrated for the target model's K geometry, so draft context creation fails:

• with default f16 draft KV: llama_init_from_model: path_kv_mean_center requires the K cache type to be Q4_0 (got f16)
• with -ctkd q4_0: load_kv_mean_center: bias tensor kv_bar.blk.3.k has 1024 elements, expected 512 (n_embd_head_k * n_head_kv)

Either way the probe reports [spec] failed to measure draft model memory and speculative decoding is silently disabled (common_speculative_init: no implementations specified) — making --kv-mean-center and -md mutually exclusive in practice.

Fix

Clear kv_mean_center_path on the copied draft params at both construction sites in tools/server/server-context.cpp (the VRAM-measurement probe and the real draft context), following the existing pattern of draft-specific overrides (cache_type_k/cache_type_v). The main context keeps the bias.

Coverage note: the only consumption site is common_context_params_to_llama (common/common.cpp), and these are the only two non-main-context derivations of common_params in the server for the -md path (the MTP branch shares the target model and is untouched).

Verified

RTX A5000, Ternary-Bonsai-27B Q2_0 + dspark drafter + --kv-mean-center + q4_0 KV (-ctkd/-ctvd q4_0), based on 62061f9:

• bias loads for the main model only, no load_kv_mean_center errors
• common_speculative_impl_draft_dspark: adding speculative implementation 'draft-dspark' — spec active alongside the bias
• greedy decode ~59 tok/s with drafter vs ~48 AR; BFCL simple_python 95.0% with drafter vs 95.25% without (quality unaffected)

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Prevents target-model KV mean-centering bias from being applied to incompatible draft-model contexts.

Changes:

• Clears the inherited bias path during draft VRAM measurement.
• Clears it when creating the actual draft context.
• Preserves bias behavior for the target and MTP contexts.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.