Qwen3.8-Flash-Next — IQ4_NL GGUF for Strix Halo

A quantization of Qwen3.8-Flash-Next (177 B parameters) built to run on a single AMD Ryzen AI Max+ with a Radeon 8060S (gfx1151) and 128 GB of unified memory.


Weights	IQ4_NL, 93.16 GiB across 9 shards
Parameters	176.94 B
Draft head	separate 2.6 GB Q8_0 file, shared embeddings
Trained context	262 144
Architecture	`qwen4exp` — gated delta-net recurrence, lightning-indexer sparse attention, per-layer embeddings

Files

Qwen3.8-Flash-Next-IQ4_NL-PROJFIX-*-of-00009.gguf — the target model. Point llama.cpp at shard 00001; it finds the rest.
mtp-Qwen3.8-Flash-Next-shared-Q8_0.gguf — the MTP draft head, exported with --mtp-shared-embd so it borrows the target's token embeddings, output norm and LM head instead of carrying its own copy of a 27.5 GB embedding table.

PROJFIX marks a re-export that corrects the projection tensor layout; earlier local builds without it are not interchangeable.

What it needs

This model uses architecture support and ROCm kernels that are not in upstream llama.cpp. It needs the strix-halo branch of pwilkin/llama.cpp:

bash <(curl -fsSL https://raw.githubusercontent.com/pwilkin/strix-halo/main/install-flash-next.sh)

That builds the pinned ROCr, HIP and llama.cpp revisions into your home directory and installs a launcher with the right flags. Nothing is written to /opt/rocm and nothing needs root.

Two flags matter more than the rest:

--load-mode none --lazy-mode on-direct

The per-layer-embedding table is 27.5 GB. on-direct serves its rows with explicit pread()s from a thread pool instead of demand-faulting them through an mmap, and --load-mode none avoids holding a second copy of every weight in page cache during load. Under unified memory with everything offloaded, the weights land in managed (anonymous) memory, so all 93 GiB already competes with the KV cache and compute buffers — without these two flags the model does not fit.

Measured throughput

Radeon 8060S, 128 GB unified, ROCm 10.0 with the retained-PM4 runtimes, -b 24576 -ub 24576 -p 16384 -n 128 -r 3:

Context depth	Prefill (t/s)	Generation (t/s)
0	1151.77 ± 8.12	24.13 ± 0.29
40 000	1059.82 ± 4.46	15.43 ± 1.77

Prefill keeps 92% of its rate at 40 000 tokens of context, which is the sparse attention path working as intended. Generation is the weaker half and is still short of what the same kernels reach out of tree; that gap is open.

A commit-by-commit account of how prefill got from 191 t/s to this, including the measurements that came out flat and the two bugs the work surfaced, is at pwilkin.github.io/strix-halo/journey.html.

Caveats

Experimental. The ROCm kernels are selected by environment variables that the installed launcher sets for you; run llama-server directly and you get the generic paths.
The 24576-token ubatch that these numbers use cannot run on upstream llama.cpp on this hardware at all: upstream's mm_ids_helper sizes its shared memory as n_tokens × 4 bytes, which is 98 KB against 64 KB of LDS on gfx1151.
Quality was checked by paired perplexity against a reference stack rather than by public benchmark scores.