A ROCmFPX GGUF quantization of DeepSeek-V4-Flash-0731, built to run the full target
locally on AMD Strix Halo (Ryzen AI MAX+ 395 / Radeon 8060S) in unified memory.
This is a baseline, deliberately. The format assignment here is uniform by tensor role
— the same public ROCmFPX recipe, applied to the new checkpoint. It exists so that our own
adaptive-format work has an honest "before" to be measured against, on the same checkpoint,
with the same harness. It is not the interesting artifact; it is the control.
What's in it
role
qtype
name
block
bits/weight
tensors
attention, dense
101
Q4_0_ROCMFP4_FAST
17 B / 32
4.25
660
down projections
104
Q3_0_ROCMFPX
14 B / 32
3.50
43
gate / up projections
107
Q2_0_ROCMFP2
10 B / 32
2.50
86
passthrough (norms, embeddings, router)
—
F32 / Q6_K
—
—
539
1328 tensors, single file, no companion sidecar required. The qtype histogram is an exact
match to the published preview-era ROCmFPX artifact, which is what makes it a fair control:
the format is held fixed and only the checkpoint differs.
What it is not
Not the Lucebox artifact.Lucebox's published
ROCMFPX GGUF is built from the
preview checkpoint. This is an independent build from 0731, matching their format.
Not adaptive. No learned codebooks, no per-expert format selection. Those land in
separate repos.
Not imatrix-calibrated. There is no calibration input at all: the assignment is fixed
by role in the export plan. If you are used to seeing quantize.imatrix.* keys in a GGUF
of this family, their absence here is correct and deliberate — see below.
Metadata provenance
This artifact was assembled using a metadata template taken from a published GGUF of the
same family, and it initially inherited that file's quantize.imatrix.* keys — which
described an imatrix calibration on someone else's build machine that played no part in
producing these weights. Those keys have been removed. What remains:
If a GGUF's metadata claims a method that did not produce it, every downstream comparison
built on it is quietly wrong. Worth checking on any quant, not just this one.
Speculative decode
Pair with the drafter for DSpark speculative decode:
DeepSeek-V4-Flash-0731-DSpark-Drafter-GGUF
— extracted from this checkpoint's integrated MTP head, so it is matched to this target.
Measurements
Not published here yet. Throughput, prefill, accept rate and quality for this baseline
and for the adaptive variants are being measured under one protocol on one box, and will be
reported together. A number measured here and compared against a number quoted from
somewhere else is not a comparison, so we would rather wait.
