Introduction of glm-moe-dsa-tiny-random-fp8-block32-rtn-format
Model Details of glm-moe-dsa-tiny-random-fp8-block32-rtn-format
GLM MoE DSA — fp8-block32: a tiny toolchain fixture
A
fp8-block32 storage/reader fixture
, built from a tiny random reference checkpoint. Use it to inspect packing, reconstruction, scope accounting, and own-head fidelity comparisons—not to choose a production quantizer.
These weights are untrained, not an assistant, and not a language-quality benchmark.
Generated text has no useful semantic quality. No upstream trained weights or training data are implied by an architecture name.
Weight lineage
This is a
quantization of a shared native checkpoint
, not an independent random initialization and not a fine-tune. HF
base_model
/
base_model_relation: quantized
metadata names the real source model.
That identity binds config and weight bytes; equal architecture names or random seeds are not used as a substitute. Quantized storage cases use round-to-nearest packing and do
not
claim GPTQ/AWQ/ModelOpt optimization or calibration.
260 tokens; independent byte tokenizer, not upstream vocabulary
CPU stack
Python 3.12; Torch 2.11.0+cpu; Transformers 5.16.1; two Torch threads in the recorded capture workflow
Remote model code
Not required: native Transformers class behind QFS
Size is serialized artifact size,
not
runtime RAM or a capacity/performance guarantee. Packed-array element counts are not model parameter counts.
What is actually included
Four text decoder layers (one dense, three MoE), MLA attention, alternating full/shared DSA indexers, 8 routed experts/top-2 plus one shared expert, and a complete untied vocabulary head. Native FP32 router buffers are retained; no MTP.
Storage intervention:
fp8-block32; random FORMAT fixture.
GLM MoE DSA; full random tiny checkpoint; header-derived GGUF geometry; not all GGUF architectures
Output-head policy:
native
. The comparison replays each side through its own sealed head, including any head error; the reference head is not substituted.
Weights reconstructed to BF16; dynamic FP8 activation quantization not captured; no serving-kernel parity claim
The repository file inventory includes
config.json
,
generation_config.json
,
tokenizer.json
,
tokenizer_config.json
,
scope.json
,
build.json
. Exact build/runtime/license inventories are linked below. A file may describe historical provenance without being an executable entry point.
Good community uses—and boundaries
Learn how to download pinned artifacts, seal a small synthetic token panel, capture hidden states and replay a full vocabulary head.
Debug model-family adapters, strict tensor loading, storage decoders, or reproducibility tooling without downloading a production-sized checkpoint.
Reproduce a narrowly scoped result, report an adapter/reader regression, and retain the source, panel and runtime identities needed to explain it.
Not established:
trained-model accuracy; useful instruction following; quantizer optimization quality; original production-weight compatibility; GPU/NPU/serving-kernel parity; cross-hardware determinism; long-context behavior outside the recorded panel; throughput or paid-compute admission.
RTN is round-to-nearest packing, not optimization.
These cases do not run GPTQ, AWQ or AutoRound optimization, calibrated ModelOpt/CT/QAT training, or activation-aware tuning. CPU reconstruction does not execute the original packed GEMM or serving activation arithmetic.
advisory
; reconstruction/activation omissions remain in force
Top-1 agreement
0.8611111111111112 on this panel only
Full comparison receipt
includes per-context/tokenwise evidence, comparison gates and disclosures. Receipt self-digest:
409c3c36bee9b60ae381787542c2c96ce8f56a1e8f113a0b9dff7a0c5204650c
. No threshold or usable determinism floor is invented here.
Material disclosures from the observed receipt:
HEAD-1d: each side replayed through its own sealed head (reference 2364995d4009, candidate 2364995d4009); head error is inside the measurement, as under HEAD-2, and nothing is substituted. The heads are content-identical.
candidate was captured from a WEIGHTS-ONLY RECONSTRUCTION (fp8-block-dequant-to-bf16, output bfloat16). The stored weights were decoded before the model forward; native quantized GEMM and serving activation arithmetic are not measured. Decoder provenance is in the sealed runtime receipt. The comparison is advisory.
candidate was captured from a bf16 materialisation of its block-scaled FP8 weights (fp8-block-dequant-to-bf16); the checkpoint declares activation_scheme: dynamic, so a served W8A8 deployment also quantizes activations per token at runtime. That term is not in this number, which is expected to understate the served divergence; it is not a mathematical bound. The comparison is advisory.
Runtime requirements and safe local reproduction
Replay the published evidence without loading a model
This uses QFS's existing
NumPy FP64 comparator in a Torch-free environment
.
It reads the stored hidden states and each side's own head. No model forward,
remote model code, GPU, upload or registry mutation is involved. Runtime receipts
name the actual backend; last-bit differences from another FP64 implementation
are not a new quality claim. Synthetic exact controls remain zero.
Set
QFS
to a reviewed Quant Fidelity Suite checkout and use Bash:
Reconstructed format comparisons are intentionally
advisory
and normally
return exit code 2 while writing a valid receipt. Inspect that receipt; do not
silence refusals or interpret an advisory result as a production-quality ranking.
Capture the actual checkpoint
Capture uses a separate pinned Torch CPU environment. The input below is the
original token-panel format
, not a sealed capture's internal
panel/
folder.
The native source, tokenizer and model revisions remain explicit. Set
AUTHOR
to your own HF handle; the dataset repository argument is attribution only and
nothing is uploaded by these commands.
Inspect
workflow.json
. This launcher executes its exact capture/verification
commands and selects the Torch-free interpreter only for the final comparison:
export OMP_NUM_THREADS=2 MKL_NUM_THREADS=2 OPENBLAS_NUM_THREADS=2
"$WORK/replay-env/bin/python" - "$WORK/workflow/workflow.json" <<'PY'
import json, subprocess, sys
workflow = json.load(open(sys.argv[1]))
for step in workflow["commands"]:
argv = list(step["argv"])
if step["step"] == "compare":
argv[0] = sys.executable
result = subprocess.run(argv)
if result.returncode:
raise SystemExit(result.returncode)
PY
The fixture repository retains its
mit
license in
LICENSE
. This covers the fixture only according to that exact text; architecture inspiration is not relicensing of upstream weights, cards, code, or configuration. Dependencies retain their own licenses (native Transformers implementation: Apache-2.0). No upstream trained tensors are claimed to be copied.
Full preserved earlier model card
, SHA-256
254f3a7d95d6c3f1f24df320f783e6e6eb790b091e35f8ce12dfac281f43847d
. This is an immutable historical record, including original build pins, measured resource tables, command logs, limitations and failed-attempt provenance. Historical publisher commands are
not
part of the local-use recipe above.
Retained reproduction tools/dependencies
and
source-inventory.json
. Source paths in recorded argv are historical execution locations; restore the dataset’s raw trees and map them to your own workspace. Do not substitute moving upstream branches or an unreviewed publisher workflow.
Selected original provenance fields (full tensor/component evidence remains in the linked inventories):
{"source":{"repository":"malaiwah/glm-moe-dsa-tiny-random-bf16","revision":"45563636ef723acfb826755493447dc40c7a0c37"},"format":"fp8-block32; random FORMAT fixture","classification":"lossy","architecture_restrictions":["GLM MoE DSA; full random tiny checkpoint; header-derived GGUF geometry; not all GGUF architectures"],"activation_omissions":["Weights reconstructed to BF16; dynamic FP8 activation quantization not captured; no serving-kernel parity claim"],"head_policy":"native"}
Additional original storage-build provenance (tensor entries remain in
build.json
):
{"all_source_tensors_preserved":true,"changed_canonical_tensors":["model.layers.0.mlp.down_proj.weight","model.layers.0.mlp.gate_proj.weight","model.layers.0.mlp.up_proj.weight","model.layers.0.self_attn.indexer.wk.weight","model.layers.0.self_attn.indexer.wq_b.weight","model.layers.0.self_attn.kv_a_proj_with_mqa.weight","model.layers.0.self_attn.kv_b_proj.weight","model.layers.0.self_attn.o_proj.weight","model.layers.0.self_attn.q_a_proj.weight","model.layers.0.self_attn.q_b_proj.weight","model.layers.1.mlp.experts.0.down_proj.weight","model.layers.1.mlp.experts.0.gate_proj.weight","model.layers.1.mlp.experts.0.up_proj.weight","model.layers.1.mlp.experts.1.down_proj.weight","model.layers.1.mlp.experts.1.gate_proj.weight","model.layers.1.mlp.experts.1.up_proj.weight","model.layers.1.mlp.experts.2.down_proj.weight","model.layers.1.mlp.experts.2.gate_proj.weight","model.layers.1.mlp.experts.2.up_proj.weight","model.layers.1.mlp.experts.3.down_proj.weight","model.layers.1.mlp.experts.3.gate_proj.weight","model.layers.1.mlp.experts.3.up_proj.weight","model.layers.1.mlp.experts.4.down_proj.weight","model.layers.1.mlp.experts.4.gate_proj.weight","model.layers.1.mlp.experts.4.up_proj.weight","model.layers.1.mlp.experts.5.down_proj.weight","model.layers.1.mlp.experts.5.gate_proj.weight","model.layers.1.mlp.experts.5.up_proj.weight","model.layers.1.mlp.experts.6.down_proj.weight","model.layers.1.mlp.experts.6.gate_proj.weight","model.layers.1.mlp.experts.6.up_proj.weight","model.layers.1.mlp.experts.7.down_proj.weight","model.layers.1.mlp.experts.7.gate_proj.weight","model.layers.1.mlp.experts.7.up_proj.weight","model.layers.1.mlp.shared_experts.down_proj.weight","model.layers.1.mlp.shared_experts.gate_proj.weight","model.layers.1.mlp.shared_experts.up_proj.weight","model.layers.1.self_attn.kv_a_proj_with_mqa.weight","model.layers.1.self_attn.kv_b_proj.weight","model.layers.1.self_attn.o_proj.weight","model.layers.1.self_attn.q_a_proj.weight","model.layers.1.self_attn.q_b_proj.weight","model.layers.2.mlp.experts.0.down_proj.weight","model.layers.2.mlp.experts.0.gate_proj.weight","model.layers.2.mlp.experts.0.up_proj.weight","model.layers.2.mlp.experts.1.down_proj.weight","model.layers.2.mlp.experts.1.gate_proj.weight","model.layers.2.mlp.experts.1.up_proj.weight","model.layers.2.mlp.experts.2.down_proj.weight","model.layers.2.mlp.experts.2.gate_proj.weight","model.layers.2.mlp.experts.2.up_proj.weight","model.layers.2.mlp.experts.3.down_proj.weight","model.layers.2.mlp.experts.3.gate_proj.weight","model.layers.2.mlp.experts.3.up_proj.weight","model.layers.2.mlp.experts.4.down_proj.weight","model.layers.2.mlp.experts.4.gate_proj.weight","model.layers.2.mlp.experts.4.up_proj.weight","model.layers.2.mlp.experts.5.down_proj.weight","model.layers.2.mlp.experts.5.gate_proj.weight","model.layers.2.mlp.experts.5.up_proj.weight","model.layers.2.mlp.experts.6.down_proj.weight","model.layers.2.mlp.experts.6.gate_proj.weight","model.layers.2.mlp.experts.6.up_proj.weight","model.layers.2.mlp.experts.7.down_proj.weight","model.layers.2.mlp.experts.7.gate_proj.weight","model.layers.2.mlp.experts.7.up_proj.weight","model.layers.2.mlp.shared_experts.down_proj.weight","model.layers.2.mlp.shared_experts.gate_proj.weight","model.layers.2.mlp.shared_experts.up_proj.weight","model.layers.2.self_attn.indexer.wk.weight","model.layers.2.self_attn.indexer.wq_b.weight","model.layers.2.self_attn.kv_a_proj_with_mqa.weight","model.layers.2.self_attn.kv_b_proj.weight","model.layers.2.self_attn.o_proj.weight","model.layers.2.self_attn.q_a_proj.weight","model.layers.2.self_attn.q_b_proj.weight","model.layers.3.mlp.experts.0.down_proj.weight","model.layers.3.mlp.experts.0.gate_proj.weight","model.layers.3.mlp.experts.0.up_proj.weight","model.layers.3.mlp.experts.1.down_proj.weight","model.layers.3.mlp.experts.1.gate_proj.weight","model.layers.3.mlp.experts.1.up_proj.weight","model.layers.3.mlp.experts.2.down_proj.weight","model.layers.3.mlp.experts.2.gate_proj.weight","model.layers.3.mlp.experts.2.up_proj.weight","model.layers.3.mlp.experts.3.down_proj.weight","model.layers.3.mlp.experts.3.gate_proj.weight","model.layers.3.mlp.experts.3.up_proj.weight","model.layers.3.mlp.experts.4.down_proj.weight","model.layers.3.mlp.experts.4.gate_proj.weight","model.layers.3.mlp.experts.4.up_proj.weight","model.layers.3.mlp.experts.5.down_proj.weight","model.layers.3.mlp.experts.5.gate_proj.weight","model.layers.3.mlp.experts.5.up_proj.weight","model.layers.3.mlp.experts.6.down_proj.weight","model.layers.3.mlp.experts.6.gate_proj.weight","model.layers.3.mlp.experts.6.up_proj.weight","model.layers.3.mlp.experts.7.down_proj.weight","model.layers.3.mlp.experts.7.gate_proj.weight","model.layers.3.mlp.experts.7.up_proj.weight","model.layers.3.mlp.shared_experts.down_proj.weight","model.layers.3.mlp.shared_experts.gate_proj.weight","model.layers.3.mlp.shared_experts.up_proj.weight","model.layers.3.self_attn.kv_a_proj_with_mqa.weight","model.layers.3.self_attn.kv_b_proj.weight","model.layers.3.self_attn.o_proj.weight","model.layers.3.self_attn.q_a_proj.weight","model.layers.3.self_attn.q_b_proj.weight"],"claim":"FORMAT fixture; weights-reconstructed advisory; activation-not-captured; no GPU validation","format":"fp8_e4m3","license":"mit","own_head":true,"quantization_config":{"activation_scheme":"dynamic","fmt":"e4m3","modules_to_not_convert":["lm_head","model.embed_tokens","model.layers.0.input_layernorm","model.layers.0.post_attention_layernorm","model.layers.0.self_attn.indexer.k_norm","model.layers.0.self_attn.indexer.weights_proj","model.layers.0.self_attn.kv_a_layernorm","model.layers.0.self_attn.q_a_layernorm","model.layers.1.input_layernorm","model.layers.1.mlp.gate","model.layers.1.post_attention_layernorm","model.layers.1.self_attn.kv_a_layernorm","model.layers.1.self_attn.q_a_layernorm","model.layers.2.input_layernorm","model.layers.2.mlp.gate","model.layers.2.post_attention_layernorm","model.layers.2.self_attn.indexer.k_norm","model.layers.2.self_attn.indexer.weights_proj","model.layers.2.self_attn.kv_a_layernorm","model.layers.2.self_attn.q_a_layernorm","model.layers.3.input_layernorm","model.layers.3.mlp.gate","model.layers.3.post_attention_layernorm","model.layers.3.self_attn.kv_a_layernorm","model.layers.3.self_attn.q_a_layernorm","model.norm"],"quant_method":"fp8","weight_block_size":[32,32]},"quantizer":"engines/tools/fp8_quantize.py","quantizer_sha256":"57dccad7d7c63ce8ca368546b2540fe37f8ae2f1d08a8e5607160e5eb5ddfc02","source_checkpoint_sha256":"9efe50a1685e157c1b99c3824094cf4fe269c7d332bdd0006145264949c5cec1","source_repository":"malaiwah/glm-moe-dsa-tiny-random-bf16","source_revision":"45563636ef723acfb826755493447dc40c7a0c37","source_tensor_count":139}
Runs of malaiwah glm-moe-dsa-tiny-random-fp8-block32-rtn-format on huggingface.co
12
Total runs
-3
24-hour runs
-6
3-day runs
-7
7-day runs
-7
30-day runs
More Information About glm-moe-dsa-tiny-random-fp8-block32-rtn-format huggingface.co Model
More glm-moe-dsa-tiny-random-fp8-block32-rtn-format license Visit here:
glm-moe-dsa-tiny-random-fp8-block32-rtn-format huggingface.co is an AI model on huggingface.co that provides glm-moe-dsa-tiny-random-fp8-block32-rtn-format's model effect (), which can be used instantly with this malaiwah glm-moe-dsa-tiny-random-fp8-block32-rtn-format model. huggingface.co supports a free trial of the glm-moe-dsa-tiny-random-fp8-block32-rtn-format model, and also provides paid use of the glm-moe-dsa-tiny-random-fp8-block32-rtn-format. Support call glm-moe-dsa-tiny-random-fp8-block32-rtn-format model through api, including Node.js, Python, http.
glm-moe-dsa-tiny-random-fp8-block32-rtn-format huggingface.co is an online trial and call api platform, which integrates glm-moe-dsa-tiny-random-fp8-block32-rtn-format's modeling effects, including api services, and provides a free online trial of glm-moe-dsa-tiny-random-fp8-block32-rtn-format, you can try glm-moe-dsa-tiny-random-fp8-block32-rtn-format online for free by clicking the link below.
malaiwah glm-moe-dsa-tiny-random-fp8-block32-rtn-format online free url in huggingface.co:
glm-moe-dsa-tiny-random-fp8-block32-rtn-format is an open source model from GitHub that offers a free installation service, and any user can find glm-moe-dsa-tiny-random-fp8-block32-rtn-format on GitHub to install. At the same time, huggingface.co provides the effect of glm-moe-dsa-tiny-random-fp8-block32-rtn-format install, users can directly use glm-moe-dsa-tiny-random-fp8-block32-rtn-format installed effect in huggingface.co for debugging and trial. It also supports api for free installation.
glm-moe-dsa-tiny-random-fp8-block32-rtn-format install url in huggingface.co:
