kingjones777/Ling-3.0-flash-base-midtrain-ROCmFP4-COHERENT-GGUF

Ling-3.0-flash-base-midtrain — ROCmFP4 for AMD Strix Halo (gfx1151)

✅ the first ROCmFP4 build of Ling-3.0-flash-base-midtrain, published with a measured MTP curve

Checked 2026-08-22 against every public GGUF of this checkpoint. The only other GGUF build (avar6/Ling-3.0-flash-base-midtrain-gguf) ships a single standard k-quant, Q5_K_M. ROCmFP4 is a runtime tensor format that exists only in the ROCmFPX fork of llama.cpp. Repository-content comparison only — no third-party build was run or benchmarked here.

A 4-bit ROCmFP4 quantisation of Ling-3.0-flash-base-midtrain for AMD Ryzen AI Max+ 395 / Radeon 8060S / gfx1151, with the multi-token-prediction (MTP) draft head preserved.

⚠️ This is a base checkpoint, not an instruct model

Ling-3.0-flash-base-midtrain is a pretrained / base checkpoint released by inclusionAI for continued pretraining, domain adaptation and fine-tuning. It is not instruction-tuned. It ships a chat_template.jinja, but that is a tokenizer asset — it does not make the weights conversational. Prompt it as a text continuation model. For chat, use inclusionAI/Ling-3.0-flash instead.

midtrain is the mid-training checkpoint — after the 30T pretraining stage and after long-context extension, but before any instruction tuning.

This checkpoint carries the full context_length = 262,144 and rope_theta = 6000000 of the long-context-extended Flash base line. (The 30T checkpoint, by contrast, declares only 8,192.)

The file

ftype	102 — Q4_0_ROCMFP4_COHERENT
size	72,123,713,664 bytes (67.17 GiB)
parameters	127.49 B (512 experts × 3.9 B, 8 active)
architecture	bailing-hybrid — hybrid KDA linear attention + MLA
tensors	938 · block_count 43 (42 layers + 1 MTP layer)
context	262,144
rope_theta	6000000

Head protection, verified in the finished file (not merely requested at quantise time, and re-audited after the metadata rename that produced the final bytes above):

output.weight        Q6_K
token_embd.weight    Q6_K
histogram: ROCmFP4(type 100) x545, F32 x390, Q6_K x2, Q8_0 x1

tie_word_embeddings is false on this model, so --output-tensor-type does real work here — the COHERENT tier on its own leaves output.weight at 4-bit. Both heads were forced to Q6_K and audited on exact tensor names.

Architecture notes

Ling-3.0-flash interleaves two attention types. head_count_kv is a per-layer array where 0 marks a KDA linear-attention layer and 1 a full MLA layer: 1 MLA layer in every 6. MLA uses a compressed KV path (kv_lora_rank 512) with a plain wide query projection (q_lora_rank: null). The blk.42 MTP layer is retained in full, including nextn.eh_proj, nextn.enorm, nextn.hnorm and nextn.shared_head_norm, with the unfused attn_k_b / attn_v_b form that the MTP path requires.

Measured throughput

AMD Ryzen AI Max+ 395, Radeon 8060S (gfx1151), ROCm 7.2.4, 128 GB unified memory. llama-cli, -dio -ngl 999 -st -c 2048 -n 512 --temp 0 --seed 1234, 3 repetitions per config, measured on an otherwise idle box.

config	flags	generation (median)	runs
no drafter	--spec-type none	36.6 t/s	36.6 / 36.6 / 36.6
MTP n-max 3	--spec-type draft-mtp --spec-draft-ngl 999 --spec-draft-n-max 3	43.0 t/s	43.2 / 43.0 / 43.0

MTP is worth +17.5% on this checkpoint, with disjoint ranges.

All three Ling-3.0-flash base checkpoints measure the same no-drafter baseline to the decimal on identical hardware and flags, which is the cross-check for this figure:

checkpoint	no drafter	MTP n-max 3	effect
Ling-3.0-flash-base	36.6 t/s	42.3 t/s	+15.6%
Ling-3.0-flash-base-30T	36.6 t/s	41.4 t/s	+13.1%
Ling-3.0-flash-base-midtrain (this file)	36.6 t/s	43.0 t/s	+17.5%

MTP is reliably positive across the whole Ling-3.0-flash base family. It is not reliable on Ling-3.0-tiny, where the same measurement gives +7.5% / +5.1% / −4.4% across the three checkpoints — the draft head is trained with the model, so its value belongs to the specific (size, checkpoint) pair rather than to the architecture. Measure before enabling it.

Requirements

This file uses the ROCmFP4 tensor format and the bailing-hybrid architecture. It requires a build of ROCmFPX that carries both. Stock llama.cpp will not load it. Verify with strings libllama.so | grep bailing-hybrid — the architecture table lives in the shared library, not in the thin CLI binary.

llama-cli -m Ling-3.0-flash-base-midtrain-Q4_0_ROCMFP4_COHERENT.gguf \
  -dio -ngl 999 -c 2048 -n 512 \
  --spec-type draft-mtp --spec-draft-ngl 999 --spec-draft-n-max 3 \
  -p "The history of mathematics begins in ancient times. One of the earliest known"

-dio (direct I/O) is recommended. At -ngl 999 the HIP backend copies offloaded tensors out of file-backed pages into device allocations, so without direct I/O the source pages and the device buffer are resident simultaneously — roughly twice the model size, which is tight on a 128 GB box.

Sample output

Continuation from "The history of mathematics begins in ancient times. One of the earliest known":

mathematical texts is the Rhind Papyrus, which dates back to around 1650 BCE in ancient Egypt. This papyrus, named after the Scottish antiquarian Henry Rhind who purchased it in 1858, contains a collection of mathematical problems and solutions that provide valuable insights into the mathematical knowledge and practices of the time. The Rhind Papyrus includes problems related to arithmeti

Not measured

• Perplexity is not published for this build. A 127 B model at this size exceeds a practical evaluation budget on a single Strix Halo box. Quality evidence here is limited to the coherence check above and the verified tensor-level audit.
• Output determinism under MTP was not tested on this checkpoint. On the sibling Ling-3.0-flash-base, MTP was found not to be output-deterministic at --temp 0 with a fixed seed. Assume the same here unless you verify it.
• n-max 5 was not swept on this checkpoint. n-max 3 is the published setting.

Provenance

Converted from inclusionAI/Ling-3.0-flash-base-midtrain at revision 34f7c1ed096bdb3118ec1474132ad21794d4510a to BF16 GGUF (938 tensors), then quantised to ftype 102 with --output-tensor-type q6_K, then general.name set to Ling-3.0-flash-base-midtrain and the heads re-audited on the finished file. Licence MIT, inherited from the base model.