Preserve pinned upstream MTP head in native-8k v0.2 GGUFs (unchanged adapters)

README.md

@@ -21,6 +21,8 @@ GGUF quantizations of a fine-tuned model for translating Japanese ASMR transcrip
 
 The model normalizes imperfect audio transcriptions, applies domain-specific glossaries, and translates character dialogue while retaining emotion and nuances.
 
+This variant preserves the upstream Qwen3.5 MTP / speculative-decoding head in GGUF format so it can be used with MTP-capable llama.cpp builds.
+
 ## Echo Mode
 
 The model echoes the source Japanese text in an `"input"` field and records applied terms in a per-entry `"glossary"` object alongside the target translation. This provides an explicit source anchor that can reduce omitted or drifted segments, but it does not guarantee immunity to long-context repetition or noisy-ASR failures.
@@ -29,10 +31,10 @@ The model echoes the source Japanese text in an `"input"` field and records appl
 
 | Quantization | Filename | Size | Description |
 |---|---|---|---|
-| q4_k_m | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf` | 5.2 GB | Good balance of quality and size |
-| q6_k | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q6_k.gguf` | 6.9 GB | Higher quality, moderate size |
-| q8_0 | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q8_0.gguf` | 8.9 GB | Near-lossless quality |
-| bf16 | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-bf16.gguf` | 16.7 GB | Full BF16, no quantization loss |
+| q4_k_m | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf` | 5.4 GB | Good balance of quality and size |
+| q6_k | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q6_k.gguf` | 7.0 GB | Higher quality, moderate size |
+| q8_0 | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q8_0.gguf` | 9.1 GB | Near-lossless quality |
+| bf16 | `asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-bf16.gguf` | 17.1 GB | Full BF16, no quantization loss |
 
 ## Prompt Example
 
@@ -98,6 +100,12 @@ llama-server -m asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf -c 8192 --port
 llama-cli -m asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf -c 8192 -p "<your prompt>" -n 2048
 ```
 
+### llama-cli with MTP
+
+```bash
+llama-cli -m asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf -c 8192 --spec-type draft-mtp --spec-draft-n-max 6 -p "<your prompt>" -n 2048
+```
+
 ## Structured Decoding (Recommended)
 
 This model outputs JSON arrays. Using structured decoding (e.g. GBNF grammar or JSON schema constraints) avoids wasted computation on malformed output and guarantees valid JSON on every generation.
@@ -203,3 +211,8 @@ Treat long-context use of this variant as **experimental**. Prefer shorter windo
 ### Content Notice
 
 The training domain includes adult ASMR dialogue and may produce sexually explicit text. This model is intended for transcription translation and subtitle-processing workflows.
+
+
+## MTP-preserving export
+
+This revision preserves and verifies all 15 upstream MTP tensors against the pinned base checkpoint after the LoRA merge. The native-8192 adapter is unchanged; no retraining was performed. The MTP head itself is not fine-tuned. Earlier revisions of this v0.2 repository omitted MTP. Filenames and translation weights remain compatible; use a fresh repository revision to avoid cached non-MTP files. All four quantizations have verified MTP metadata and tensor inventory. Q4_K_M was GPU-smoke-tested with llama.cpp b9247, context 8192, and `--spec-type draft-mtp --spec-draft-n-max 3`. This is a runtime smoke test, not a new quality evaluation or a speedup guarantee. See `mtp-preservation.json` and `mtp-release-verification.json` for provenance and checks.

asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-bf16.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:d195f732d70e053e84ed48fbbafbe00bf006e258a324a15cdde4186241e5dda4
-size 17920696864
+oid sha256:b2fd49a58f5f829276ba4da8715875113064b45dc971bb403db3bbe4284e54d9
+size 18407321056

asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a867434b5bde577e170582624a0ed0d178f56c16c07712e30b780fa9bad25105
-size 5629108768
+oid sha256:e185134cb504cbb93e42c9da03461331b063155bcec3530daa87bc0e610fee77
+size 5780090336

asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q6_k.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:c1a98bdac4fab31bf109228c75376ccff3cfb9201a10e745845d713dce4801ff
-size 7359259168
+oid sha256:4bcab83434cf001ac7da9a98e2ba4d4809e735aa2f02a71df4efb3408244ea3f
+size 7558901216

asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q8_0.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:62a1e33b8c78360259d6d9f3f6be4049861551c4e98b328dcc9586d90a263050
-size 9527501344
+oid sha256:49a46cb38baee3f3b87814198a6a6da41b120c208de560ce02b6e9ae6c43ec5b
+size 9786060256

mtp-preservation.json

@@ -0,0 +1,23 @@
+{
+  "base_revision": "005429cee5cb648998cf2b70eebdd83175989c9a",
+  "base_model": "unsloth/Qwen3.5-9B",
+  "restored": [],
+  "tensor_sha256": {
+    "mtp.layers.0.input_layernorm.weight": "c8913bfe7cef186fb59b7f1ca80d391ba5eaec45c9c8db54f54e63f6410ae633",
+    "mtp.layers.0.post_attention_layernorm.weight": "63500aa0ed93918f6f3d5f1bfd1656921543931392a06267f81c0b3ad680fdcf",
+    "mtp.layers.0.self_attn.k_norm.weight": "8f5a8071649fa9a7b6b6afcea6d5404cd849640be2c4f90e6759c4c0a60120ac",
+    "mtp.layers.0.self_attn.k_proj.weight": "1b0129a3c6f4add6c8417d67e84050f8595f96f7693511beaace9ef56d1ad30a",
+    "mtp.layers.0.self_attn.o_proj.weight": "31140d974d94190127d0f3d908a18bafe5fba0d8c3d7f522e012013795c89924",
+    "mtp.layers.0.self_attn.q_norm.weight": "10e6c9fa42ceb72373c207422c305663a89dfd08f5332cbbe6d1057301fe57d1",
+    "mtp.layers.0.self_attn.v_proj.weight": "c21332947c44203ca197d014de23559bf6e88d7c0debca8d4ecad735d717e2c9",
+    "mtp.norm.weight": "7e4daf06ad25b834b3b95c9592fa690b75a4a90fb9f4128ccee8537d7f15988e",
+    "mtp.pre_fc_norm_embedding.weight": "8b9bca62497def4783b20dfd58ddf32573a255b44a79f0acf8ca9c611b7bea15",
+    "mtp.pre_fc_norm_hidden.weight": "a7410736f9962dd6ebd8ac2b4c355898d057296a03c47795223b7298cac8988c",
+    "mtp.fc.weight": "a4639d8f4b81cbdc65c61f1cba82816ae1534ec494a95a0b677dba98f04f4017",
+    "mtp.layers.0.self_attn.q_proj.weight": "17d3aac05cb017e9ef98f93fe567f1c1ffae08fbe5001b80e67829bda189a584",
+    "mtp.layers.0.mlp.down_proj.weight": "52c72564f7da59c25233b2194a79239cc1e69dd6694130aae87d5fffc472707c",
+    "mtp.layers.0.mlp.gate_proj.weight": "dd3e6d05e9c519ebb16c5eee64b4d4d217d7efb8ae7c88e8dfa96f6e6f7d3eac",
+    "mtp.layers.0.mlp.up_proj.weight": "79e02e92d4cb775120f48cd523577345b277dacb749e56e2d52532583f92800e"
+  },
+  "mtp_layers": 1
+}

mtp-release-verification.json

@@ -0,0 +1,148 @@
+{
+  "adapter_sha256": "8000e9e54726e1943fd0c63c7f1249a351f462c2902274539e9245f483443740",
+  "files": {
+    "q4_k_m": {
+      "sha256": "e185134cb504cbb93e42c9da03461331b063155bcec3530daa87bc0e610fee77",
+      "size": 5780090336,
+      "metadata": {
+        "qwen35.block_count": 33,
+        "qwen35.nextn_predict_layers": 1
+      },
+      "mtp_tensors": [
+        "blk.32.attn_k.weight",
+        "blk.32.attn_k_norm.weight",
+        "blk.32.attn_norm.weight",
+        "blk.32.attn_output.weight",
+        "blk.32.attn_q.weight",
+        "blk.32.attn_q_norm.weight",
+        "blk.32.attn_v.weight",
+        "blk.32.ffn_down.weight",
+        "blk.32.ffn_gate.weight",
+        "blk.32.ffn_up.weight",
+        "blk.32.nextn.eh_proj.weight",
+        "blk.32.nextn.enorm.weight",
+        "blk.32.nextn.hnorm.weight",
+        "blk.32.nextn.shared_head_norm.weight",
+        "blk.32.post_attention_norm.weight"
+      ],
+      "tensor_count": 442
+    },
+    "q6_k": {
+      "sha256": "4bcab83434cf001ac7da9a98e2ba4d4809e735aa2f02a71df4efb3408244ea3f",
+      "size": 7558901216,
+      "metadata": {
+        "qwen35.block_count": 33,
+        "qwen35.nextn_predict_layers": 1
+      },
+      "mtp_tensors": [
+        "blk.32.attn_k.weight",
+        "blk.32.attn_k_norm.weight",
+        "blk.32.attn_norm.weight",
+        "blk.32.attn_output.weight",
+        "blk.32.attn_q.weight",
+        "blk.32.attn_q_norm.weight",
+        "blk.32.attn_v.weight",
+        "blk.32.ffn_down.weight",
+        "blk.32.ffn_gate.weight",
+        "blk.32.ffn_up.weight",
+        "blk.32.nextn.eh_proj.weight",
+        "blk.32.nextn.enorm.weight",
+        "blk.32.nextn.hnorm.weight",
+        "blk.32.nextn.shared_head_norm.weight",
+        "blk.32.post_attention_norm.weight"
+      ],
+      "tensor_count": 442
+    },
+    "q8_0": {
+      "sha256": "49a46cb38baee3f3b87814198a6a6da41b120c208de560ce02b6e9ae6c43ec5b",
+      "size": 9786060256,
+      "metadata": {
+        "qwen35.block_count": 33,
+        "qwen35.nextn_predict_layers": 1
+      },
+      "mtp_tensors": [
+        "blk.32.attn_k.weight",
+        "blk.32.attn_k_norm.weight",
+        "blk.32.attn_norm.weight",
+        "blk.32.attn_output.weight",
+        "blk.32.attn_q.weight",
+        "blk.32.attn_q_norm.weight",
+        "blk.32.attn_v.weight",
+        "blk.32.ffn_down.weight",
+        "blk.32.ffn_gate.weight",
+        "blk.32.ffn_up.weight",
+        "blk.32.nextn.eh_proj.weight",
+        "blk.32.nextn.enorm.weight",
+        "blk.32.nextn.hnorm.weight",
+        "blk.32.nextn.shared_head_norm.weight",
+        "blk.32.post_attention_norm.weight"
+      ],
+      "tensor_count": 442
+    },
+    "bf16": {
+      "sha256": "b2fd49a58f5f829276ba4da8715875113064b45dc971bb403db3bbe4284e54d9",
+      "size": 18407321056,
+      "metadata": {
+        "qwen35.block_count": 33,
+        "qwen35.nextn_predict_layers": 1
+      },
+      "mtp_tensors": [
+        "blk.32.ffn_down.weight",
+        "blk.32.ffn_gate.weight",
+        "blk.32.ffn_up.weight",
+        "blk.32.nextn.eh_proj.weight",
+        "blk.32.attn_q.weight",
+        "blk.32.attn_norm.weight",
+        "blk.32.post_attention_norm.weight",
+        "blk.32.attn_k_norm.weight",
+        "blk.32.attn_k.weight",
+        "blk.32.attn_output.weight",
+        "blk.32.attn_q_norm.weight",
+        "blk.32.attn_v.weight",
+        "blk.32.nextn.shared_head_norm.weight",
+        "blk.32.nextn.enorm.weight",
+        "blk.32.nextn.hnorm.weight"
+      ],
+      "tensor_count": 442
+    }
+  },
+  "runtime": {
+    "binary": "/root/llama-b9247-rocm/llama-server",
+    "command": [
+      "/root/llama-b9247-rocm/llama-server",
+      "-m",
+      "/root/asmr-one-dump/train/publish-mtp-v0.2/asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2/q4_k_m/asmr-qwen3.5-9b-zh-cn-echo-gguf-v0.2-q4_k_m.gguf",
+      "-ngl",
+      "99",
+      "-c",
+      "8192",
+      "-np",
+      "1",
+      "--port",
+      "18942",
+      "--host",
+      "127.0.0.1",
+      "--spec-type",
+      "draft-mtp",
+      "--spec-draft-n-max",
+      "3",
+      "--no-warmup",
+      "--verbose"
+    ],
+    "log_sha256": "4824e0025b2f74c1cf1af460e9f6e9eb2ada0bed0d7f82b2cf4184c98365638d",
+    "response_sha256": "cff93c67da9ee3e444356c07866c107106e601adbd4de1d0b61e37cd27a77e07",
+    "timings": {
+      "cache_n": 0,
+      "prompt_n": 24,
+      "prompt_ms": 378.189,
+      "prompt_per_token_ms": 15.757875,
+      "prompt_per_second": 63.460333325400796,
+      "predicted_n": 8,
+      "predicted_ms": 810.715,
+      "predicted_per_token_ms": 101.339375,
+      "predicted_per_second": 9.86783271556589,
+      "draft_n": 12,
+      "draft_n_accepted": 5
+    }
+  }
+}