DeepSeek-V4-Flash-Acti-MTP-W4A16-FP8 / MTP_QUANT_MANIFEST.json
LordNeel's picture
Add measured MTP draft-acceptance metrics (77.7% @128k , 84.9% @262k prod, 93-96% code) + acceptance chart; update manifest
f561982 verified
Raw
History Blame Contribute Delete
1.4 kB
{
"label": "gptq_mtp_routed_experts_v2",
"is_final_quality_preserving": true,
"method": "Frantar-style GPTQ with Cholesky H^-1; per-expert per-group sym INT4 group=128",
"calibration": {
"tokens_processed": 473372,
"expert_token_count_min": 430,
"expert_token_count_max": 175375,
"expert_token_count_mean": 11094.65625,
"n_dump_files": 17701,
"damp_fraction": 0.01
},
"simplification_note": "MTP forward replay skipped attention (treated post-attn as identity-residual input); RMSNorms, h_proj/e_proj, gate, and per-expert MLP path are real.",
"measured_draft_acceptance": {
"note": "Weighted draft acceptance from live vLLM speculative-decoding metrics (num_speculative_tokens=1). Output is verifier-exact regardless of acceptance; acceptance affects decode speed only.",
"num_speculative_tokens": 1,
"profiles": {
"general_chat_128k": {"accepted": 2354, "drafted": 3030, "weighted_acceptance": 0.777, "mean_accepted_length": 1.78, "windows": 5},
"production_262k": {"accepted": 15274, "drafted": 17991, "weighted_acceptance": 0.849, "mean_accepted_length": 1.85, "windows": 25},
"long_context_262k_research_variant": {"accepted": 319, "drafted": 357, "weighted_acceptance": 0.894, "mean_accepted_length": 1.89, "windows": 3}
},
"code_heavy_ext10_gate_adapter_assisted": {"weighted_acceptance_range": [0.93, 0.96]}
}
}