[!WARNING]
This is a custom GGUF for Ember, an
ROCmFPx-aware DeepSeek-V4 runtime for AMD Strix Halo (gfx1151). It uses
custom tensor types that mainline llama.cpp, LM Studio, and Ollama do not
implement. Loading it with an incompatible runtime may fail or produce invalid
output.
An abliterated, importance-matrix-calibrated quantization of the official
deepseek-ai/DeepSeek-V4-Flash-0731 release, plus its matching DSpark drafter.
[!IMPORTANT]
This is the text-only model. Ember's default download is now the vision-capable
DeepSeek-V4-Flash-Vision-Exp,
which also serves text and ships an image-calibrated importance matrix.
This artifact is not deprecated and has not been deleted. Set
EMBER_MODEL_REPO and EMBER_MODEL_REVISION to keep serving it. Its DSpark
drafter is the one the vision release ships with, because it measured better
than the vision checkpoint's own.
The two models' text quality has not been compared, so neither supersedes the
other on evidence.
Ember performance benchmark
Measured 2026-08-22 on one AMD Ryzen AI Max+ 395 with Radeon 8060S (gfx1151)
and 128 GB unified memory, against the artifacts published here
(target SHA-256 a936e0a5…, drafter 1a01c80e…, both re-verified against the
local copies before the run). Runtime was Ember
release 2026.8.22 in ember-rocm:7.14, with speculative decoding enabled and
per-token instrumentation disabled.
[!IMPORTANT]
These measure the production serving configuration: sparse prefill plus DSpark
speculative decode, one request at a time. They are not exact-prefill reference
numbers and they do not measure quality against BF16.
Ember sparse-prefill throughput by context length
Evaluated prompt tokens
Samples
Median prefill tok/s
Observed range
Median wall time
154
3
209.6
206.4–210.9
0.736 s
538
3
319.2
313.2–320.5
1.686 s
2,074
3
386.8
365.4–391.1
5.364 s
8,218
3
337.2
331.6–337.9
24.370 s
16,410
2
310.7
310.0–311.4
52.812 s
32,794
2
282.3
282.0–282.6
116.165 s
Every scaling prompt reported zero restored prefix tokens, so each pass evaluated
the full prompt. First runs at each shape are included in the medians.
Ember DSpark decode throughput
Three greedy 256-token samples produced 34.16 tok/s median decode
(33.67–34.16) at 0.989 median DSpark acceptance — one of the three samples
reached 1.000 — with speculation active on 3/3.
This release carries two gfx1151 kernel changes: a branch-free UE4M3 scale
decode and per-translation-unit selection of WGP vs CU mode. Against the
previous published figure of 33.60 tok/s (2026-08-10, commit 8872221,
same two model files), that is +1.7% median decode.
[!NOTE]
Acceptance is workload-dependent. A greedy 256-token continuation is highly
predictable and accepts nearly everything; separate mixed free-form sampling at
temperature 0.6 measured ~0.52 mean acceptance on the same pair. Both are real
and describe different workloads. The kernel changes in this release are
numerically inert: the differential validator reports the autoregressive,
DSpark, disk and resident-batch paths all token-exact.
The full benchmark bundle contains the
harness, raw JSONL, summary and environment JSON, both SVG sources, and the
exact model and runtime identifiers. The
2026-08-10 bundle is retained for
comparison.
Quantization quality — not yet characterized
No perplexity, no KL-divergence against BF16, no capability benchmark has been
run on this artifact. The numbers above are throughput only. The abliteration
has been validated behaviourally (see below), and the tensor layout has been
validated structurally, but the quality cost of the 2.58 BPW quantization itself
is currently unmeasured. Treat it as unknown rather than as acceptable.
Artifacts
Verify any local copy against the SHA-256 values below.
Target + matching drafter is 102,444,355,040 bytes (95.41 GiB) resident.
The abliteration band
The edit is restricted to the band where the refusal direction actually
separates — blocks 10 through 42, plus the 3 DSpark stages, leaving the first
ten blocks untouched.
edited tensors
36 (blocks 10–42 + 3 MTP)
strength (lambda)
3.5
SRA rank
0 (none)
row-norm preserve
yes
direction
rank-1, published, SHA-256 pinned
Blocks 0–9 are left untouched. Early layers carry general features and show the
weakest harmful/harmless separation, so projecting there costs capability
without adding bypass. Abliterating all 46 attention output-B tensors instead of
36 is known to damage agentic coherence: public tuning work on this model family
found the all-layers variant reaches full refusal bypass but degrades
instruction following, with the characteristic symptom being the model emitting
its skill catalog in response to a bare greeting.
The refusal direction is the published rank-1 vector, SHA-256
6e4d8a8f3aa9e21795faab2c5b14d29b019acdf2ddbfbd8238430458a5837fe0
(MIT, credit drowzeys/keys), rather than a locally extracted per-layer set.
Per-tensor relative-Frobenius edit magnitude: min 0.0494, median 0.0566,
max 0.0668.
Verification of the abliteration
Each block's attn_output_b was probed directly in the shipped weights by
measuring the refusal-direction component against an isotropic baseline
(||r^T W|| / (||W||_F / sqrt(n_out))). An untouched tensor sits near 1.0.
The bypass set is disjoint from the prompts used to derive the refusal
direction. The greeting-versus-skill-catalog case — the symptom that motivated
this build — passes. The single coherence failure is a verbosity bound: the
model answered correctly but wrapped the answer in a code fence, exceeding a
word cap.
This is a self-consistent harness, comparable across builds from this project
only. It is not a sealed judge suite and it does not measure capability.
Structural validation
The target was checked against a per-tensor type contract extracted from a GGUF
known to run on the target runtime: architecture, tensor count, block range,
pinned metadata, and the type of every tensor class.
The drafter is a separate GGUF, not part of the 43-block target: architecture
deepseek4-dflash-draft, 3 blocks, each a full 256-expert MoE block. It is
attached at runtime and the two files are never merged.
[!NOTE]
Pairing the abliterated drafter is optional. It was measured against the
stock drafter on the same target, 16 prompts (8 benign / 8 refusal-triggering),
greedy:
drafter
spec runs
mean acceptance
benign
harmful
abliterated (matched)
16/16
0.520
0.605
0.435
stock (mismatched)
16/16
0.508
0.578
0.438
The two are indistinguishable, including on refusal-triggering prompts where a
mismatch was predicted to hurt most. No throughput benefit is claimed for the
matched drafter. It is published because a drafter that models the same
distribution as its target is the defensible default, not because it measured
faster. A stock DSpark drafter works.
Speculative decoding is disabled above DFLASH_DS4_SPEC_MAX_CTX.
Note also that the drafter is not interchangeable with a ds4 --mtp DSpark
support GGUF: that artifact keeps mtp.* tensor names, while this runtime
resolves only blk.*, dflash.* and output_*. They share no tensor names.
Quantization recipe
The base preset is Q4_0_ROCMFP4_STRIX_LEAN, adapted for DeepSeek-V4's fused
MLA tensor names:
all 129 routed-expert gate/up/down tensors: affine Q2_0_ROCMFP2
(custom type 107, 2.5 BPW)
all 43 fused attn_kv tensors: dual-scale Q4_0_ROCMFP4
(custom type 100)
574 remaining dense, attention, shared-expert, indexer, and output tensors:
Q4_0_ROCMFP4_FAST (custom type 101)
all 43 attn_output_b tensors: Q8_0
token embedding: Q6_K
norms, biases, scales, and sinks: F32/I32 as appropriate
The affine expert format decodes each block as:
value = code * scale - offset, where code is one of {0, 1, 2, 3}
Its packed blocks remain 2.5 BPW while using both stored scale bytes as an
affine scale and offset rather than two symmetric half-block scales.
0731-specific importance matrix
The routed-MoE matrix was regenerated against the official 0731 weights by
collecting expert activations on ROCm:
The GGUF embeds the matrix filename, dataset identifier, chunk count, and
entry count in its quantization metadata. The downloadable .dat uses the
legacy llama.cpp importance-matrix binary format accepted by the ROCmFPx
quantizer used for this build.
Runtime validation
Checks this recipe passes end to end:
header and all 1,328 tensor descriptors parsed successfully
every special tensor matched the intended type; no missing or unexpected
overrides
final tensor end offset exactly equals the physical file size
the custom Ember ROCm runtime loaded all tensors into an 86,886.9 MiB GPU
buffer on gfx1151
health, ordinary chat completion, structured tool-call generation, and a
tool-result continuation all passed
DSpark speculative decoding initialized and ran during the smoke tests
The benchmark figures above were measured on this artifact and identify the
exact runtime, sampler, and drafter configuration.
Runtime
Built for and tested with Ember
(ember-dflash). Any runtime used instead must implement:
the DeepSeek-V4 (deepseek4) graph
custom ROCmFPx types 100, 101, and 107
the affine type-107 CPU reference path and ROCm kernels
the official DeepSeek-V4-0731 chat template, including DSML tool calls
The model runs without speculative decoding. To use the separate DSpark
drafter, the tested runtime configuration was:
