# haloq38flash
**qwen3.8-flash-next on amd strix halo โ 91g quant, 56 tok/s, 262k context**
[](https://huggingface.co/julianmb/Qwen3.8-Flash-Next-IQ4_XS-GGUF)
[](#results)
[](#results)
[](#the-converter-bug)
[](#docker)
*every published quant byte-traced back to the official checkpoint*
---
## results
91g quant ยท vulkan/radv ยท mtp sidecar ยท q8_0 kv ยท `-ub 2048` ยท temp 0 ยท 128g strix halo
| depth | plain pp / tg | mtp pp / tg |
|------:|:-----------:|:---------:|
| 0 | 92.5 / 29.9 | 87.0 / **53.1** |
| 8k | 480 / 24.1 | 458 / **56.4** |
| 32k | 397 / 20.1 | 379 / **30.2** |
| 128k | 222 / 11.0 | 214 / 18.6 |
| 256k | 139 / 6.2 | โ |
> [!NOTE]
> no collapse through 32k. the 128k+ falloff is context-mechanics
> (sparse-attention indexer), not quant size โ see the reversal below.
the 128k reversal โ the PLE quant loses under MTP at depth
at โค32k the PLE quant wins everywhere. at 128k under mtp it *loses* to the
static 116g (18.6 vs 26.9 t/s). plausible mechanism: iq4_nl noise in the
n-gram table compounds over deep history and lowers draft acceptance.
single runs, n=1 caveat. pick your file by use case โ see the table above.
---
## ๐ฆ published quants
[huggingface.co/julianmb/Qwen3.8-Flash-Next-IQ4_XS-GGUF](https://huggingface.co/julianmb/Qwen3.8-Flash-Next-IQ4_XS-GGUF)
| file | size | pick it when |
|------|------|:------------:|
| `...-IQ4_XS-`**`PLE`**`.gguf` | 91 giB | ctx โค 32k โ wins everywhere, mtp to 56 t/s |
| `...-IQ4_XS.gguf` | 116 giB | ctx โฅ 128k โ faster mtp at depth, wider fork compat |
| `mtp-...-Q8_0.gguf` | 3.9 giB | mtp sidecar, required for the speed numbers |
the PLE cut โ why the 51b n-gram table tolerates 4-bit
the PLE table is gathered 16 random rows per token via hash lookup โ there is
no matmul on the table itself, and no two consecutive tokens hit the same rows.
the rows tolerate iq4_nl (4.25 bpw) with no measurable degradation across the
depth sweep. the cut: `--tensor-type "per_layer_token_embd=IQ4_XS"` on our
quantizer โ 54g โ 27g.
**fork caveat:** engines that feed gathered PLE rows straight into mul_mat as
quantized B operands assert (ggml-vulkan.cpp:7794). verified working on the
packaged engine and rocmfpx.
### pick your setup
| your use case | quant | ctx | expect |
|---|---|---|---|
| coding agents, chat | 91g PLE | โค 32k | 56 t/s |
| long documents | 116g static | 128k | 27 t/s |
| full rag / research | 116g static + ssd streaming | 262k | 14 t/s |
---
## ๐ the converter bug
our first quant printed deterministic garbage at temp 0. bisect to root cause:
- experts, gdn reorder, ple scale, metadata: all innocent
- **97 of 388 f32 tensors differed by exactly 1.0** โ every hyper-connection
norm shipped raw where the runtime expects `raw + 1`
- cause: the checkpoint nests hyper-connections under
`attn_hyper_connection` / `mlp_hyper_connection` / `hyper_connection_mixer`,
and those names hit early-return branches in the converter that bypass the
generic `norm.weight โ +1` rule
> [!WARNING]
> **any fork rolling its own qwen4exp converter must fold `(1 + w)` into the
> hyper-connection gammas.** upstream runtime documents the contract at
> `qwen4exp.cpp:231 โ "the converter folded each gamma to (1 + w)"`. miss it
> and every layer normalizes wrong โ garbage from layer 0, all shapes correct,
> all shape-only tests pass.
fix + regression test: rocmfpx `port-qwen4exp` commit `61b6a3b48`
([pr charlie12345/ROCmFPX#98](https://github.com/charlie12345/ROCmFPX/pull/98))
---
## ๐ณ docker
```bash
git clone https://github.com/julianmb/haloq38flash && cd haloq38flash
docker compose up --build
# serve on :8080 โ vulkan/radv, no rocm install needed
```
the packaged engine is tuned for strix halo: vulkan fa/mmq kernels, graph
reuse, lazy ple streaming, quantized-kv attention โ the combination behind
the 56 t/s numbers. no manual build, no host rocm install.
add the mtp sidecar for speculative decoding:
```bash
docker compose run qwen38-flash-next /app/llama-server \
-md /models/mtp-Qwen3.8-Flash-Next-Q8_0.gguf \
--spec-type draft-mtp --spec-draft-n-max 6 --spec-draft-p-min 0.75
```
### 262k context (ssd streaming)
swap the model to the static 116g and enable lazy ple โ the n-gram table
stays on ssd (~2.5g resident), leaving room for the full context window:
```bash
docker compose run qwen38-flash-next /app/llama-server \
-m /models/Qwen3.8-Flash-Next-IQ4_XS.gguf \
-c 262144 -lm mmap --tensor-read-lazy on \
-ngl 999 -fa on -ctk q8_0 -ctv q8_0 -ub 2048 -t 4
```
---
## ๐ฌ the n-gram table at 4-bit โ what we found
the 51b ple lookup table tolerates iq4_nl (4.25 bpw) with no quality loss
across the depth sweep. but there's a depth-dependent reversal: under mtp at
128k+, the ple quant *loses* to the static quant (18.6 vs 26.9 t/s) โ the
iq4_nl noise compounds over deep n-gram history and lowers draft acceptance.
n=1, single runs. pick your file by use case.
---
## ๐ layout
| path | what |
|------|------|
| `models/` | symlink farm to local ssd (never in git) |
| `docs/` | engine merge plan, cherry-pick classification |
| `Dockerfile` | two-stage: vulkan engine build + slim runtime |
| `docker-compose.yml` | one-liner serving with recommended flags |
---
## โ ๏ธ operational gotchas (128g strix halo)
- always `-c 8192`-bounded ctx + `timeout` + `/usr/bin/time -v` โ the gguf
default 262144 + full offload hard-hung this box once
- `vm.dirty_ratio=15 / dirty_background_ratio=5` โ the 191g ple conversion
memmap wedges `balance_dirty_pages` for hours at kernel defaults
- conversion peak: ple scratch (191g) + f16 output (354g) coexist โ budget
~560g free
- `pkill -x llama-cli`, never `-f` (matches your own wrapper shell)
- gpu memory is shared with everything else on the apu โ two engines cannot
hold ~90g+ models simultaneously without an oom cascade
---
license: [qwen community license 1.0](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE) ยท base model: [Qwen/Qwen3.8-Flash-Next](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)