GLiDE-27B โ€” ROCmFP4 for AMD Strix Halo (gfx1151)

A ROCmFP4 STRIX quant of fastino/GLiDE-27B (Fastino Community License; a modified Qwen3.8-27B). Importance-matrix weighted; token embedding at Q6_K (embeddings are tied; there is no separate output.weight). Two Super ROC 4-bit prefill route files, rounded from this model's own BF16 weights, ship next to the GGUF.

Upstream GLiDE is a decision model: a 255-way readout scores typed questions in one forward pass. This GGUF is the text transformer only. The vision encoder and the decision readout from the upstream repo are not in this file. Native context in the GGUF is 262144; the serve command below uses 131072. There are no in-file nextn / MTP tensors. Serve it with the Qwen3.8-27B DFlash2 drafter.

Files

File Bytes sha256 Role
GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.gguf 14084639424 65297b09db336c35975cf696e786f51633c8c4e4736feb638a2be31a708ac504 model, recipe STRIX (ggml type 105), 4.39 BPW
GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.route.a4w 8640268824 747666a48996981685a57a5245bfa70e55da4d63d2b96f8e18a6d6fc5a278a95 4-bit prefill route (ffn_gate,ffn_up,attn_gate,attn_q,attn_qkv), clip 3, fine 64
GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.down-rot256.a4r 2878866208 b90bea9a0a91b612539c0484911cd835fbe7c622b64788f62fa87fca4dd114fb rotated ffn_down route, block 256, clip 3, fine 64

The route files belong to this GGUF only. The engine checks layer count and width, so using them with any other 27B file still loads and gives wrong answers.

Drafter (not in this repository): Qwen3.8-27B-DFlash2-Q4_K_M.gguf from z-lab/Qwen3.8-27B-DFlash2-GGUF.

How it was built

  • Source: fastino/GLiDE-27B BF16 safetensors, converted to a single BF16 GGUF (qwen35, 25,624,600,064 parameters, 850 tensors, 64 blocks, vocab 248320, tied embeddings). Visual tensors and the 255-way readout were not written into the GGUF.
  • Importance matrix: 531 chunks ร— 512 tokens of the same agentic calibration text used for the Qwen3.8-27B STRIX file, computed on a Q8_0 of GLiDE (496 imatrix entries).
  • Quantize: llama-quantize --imatrix <that file> --output-tensor-type q6_K --token-embedding-type q6_K <BF16> <out> Q4_0_ROCMFP4_STRIX 16. Census: 850 tensors; token_embd.weight converted to Q6_K; model 48880.09 MiB (16.00 BPW) โ†’ 13421.67 MiB (4.39 BPW). Wall 188.5 s. Tensor types in the STRIX file: 416 ร— ggml type 101, 80 ร— type 100, 353 ร— F32, 1 ร— Q6_K.
  • Prefill route: Super ROC int4_weights from the GLiDE BF16 GGUF. Column scales dumped on this STRIX file (8K-token prompt, context 4096). Rotated-down scales dumped on the held-out quality text with the five-tensor route on and SUPERROC_INT4_DOWN_ROT=256.

Requires a runtime that loads ROCmFP4 (ggml types 100โ€“106). Stock llama.cpp rejects these tensor types.

Run it

SUPERROC_ATTN_FOLDED=1 SUPERROC_ATTN_VW=1 \
SUPERROC_INT4_WEIGHTS=/path/to/GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.route.a4w \
SUPERROC_INT4_TENSORS=ffn_gate,ffn_up,attn_gate,attn_q,attn_qkv \
SUPERROC_INT4_DOWN_WEIGHTS=/path/to/GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.down-rot256.a4r \
SUPERROC_INT4_DOWN_ROT=256 \
superroc serve llm \
  -m /path/to/GLiDE-27B-Q4_0_ROCMFP4_STRIX-imatrix-headQ6K.gguf \
  --served-model-name glide-27b \
  --context 131072 --think off -t 0 --presence-penalty 0 \
  --speculative dflash2 --dflash-model /path/to/Qwen3.8-27B-DFlash2-Q4_K_M.gguf \
  --prompt-lookup

ROCm 7.2.4, LD_LIBRARY_PATH pointing at that ROCm lib. Thinking off. Presence penalty 0.

This GGUF is a next-token server of the text backbone. It does not expose GLiDE's native decision API.

Quality

Perplexity with Super ROC's qwen35_ppl on a held-out text, context 4096, 5 chunks, 10,235 scored tokens (second half of each chunk). Corpus sha256 167d6b72bcaf3e0139fa407889b7b1aeee18e304ec5b5ce6135efb669b9647f8. Lower is better. GLiDE's native head is a 255-way decision readout, not a generative lm_head, so absolute generative PPL is large; the useful figure is the ratio to this file's own Q8_0.

File / route PPL vs Q8_0
Q8_0 reference (not shipped) 373673 1.000
STRIX, five-tensor route on 385580 1.0319
STRIX, five-tensor route + rotated down + folded attention VW 380893 1.0193

The 1.02 ร— Q8_0 bar is 381146.46. The full stack (the serve command above) is 1.0193 ร— Q8_0.

Chat completions on this GGUF return HTTP 200 with finish_reason=length; the decoded text is not coherent English (no decision readout, no generative training objective). A weather-tool request also returned HTTP 200 with finish_reason=length and no tool_calls. Second turn of an 1,802-token prefix reused 1,802 / 1,802 cached tokens; the follow-up turn reused 1,802 / 1,899.

Speed

AMD Ryzen AI Max+ 395 (Radeon 8060S, gfx1151), 128 GB unified, ROCm 7.2.4. Super ROC, the serve command above, temperature 0, cold prompt cache (cached_tokens=0 on every timed request), means of two runs.

Measurement Result
Prompt processing, 1,566-token prompt 455.9 tok/s
Generation after that prompt (up to 256 tokens) 17.7 tok/s
Draft tokens accepted on that prompt 75 / 263
Prompt processing, 13,384-token prompt 820.85 tok/s
Generation after that prompt 20.05 tok/s
Draft tokens accepted on that prompt 99 / 212
Prompt processing, 26,821-token prompt 756.45 tok/s
Generation after that prompt 18.0 tok/s
Draft tokens accepted on that prompt 90 / 233
Code generation, 6 prompts, up to 256 tokens, mean of two runs (range) 20.73 tok/s (19.9โ€“22.4)
Draft tokens accepted on those code prompts 1112 / 2554
Tool calls valid, tool_choice required / auto 0/8 / 0/8

The DFlash2 drafter is the base Qwen3.8-27B file. Generation on this GGUF fills the 256-token cap; the table is the measured acceptance on this file.

Architecture

GLiDE-27B's text transformer is dense hybrid-attention qwen35: 64 layers, hidden 5120, FFN 17408, SSM inner 6144, vocab 248320, tied embeddings. 850 tensors. Native context in the GGUF is 262144; the serve command above uses 131072.

License

Base model fastino/GLiDE-27B, Fastino Community License (see LICENSE in this repository). GLiDE is a modified Qwen3.8-27B; Qwen3.8-27B remains Apache-2.0 (licenses/Qwen3.8-27B-LICENSE). This repository is the quantised weights, the two route files, and the measurements above. A quantized copy is a Derivative Work under the Fastino Community License: include this license with the files; commercial production use above $10,000,000 annual revenue needs a separate written agreement; commercial use requires the attribution "Built with Fastino GLiDE".

Downloads last month
140
GGUF
Model size
26B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

4-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for kingjones777/GLiDE-27B-ROCmFP4-STRIX-GGUF

Quantized
(6)
this model