thaki-AI commited on
Commit
1c1dda6
·
verified ·
1 Parent(s): 9cffdd1

D81 aggregate results (private staging)

Browse files
MANIFEST.sha256 ADDED
@@ -0,0 +1,8 @@
 
 
 
 
 
 
 
 
 
1
+ 122535d420be427183233299ff57adadf7610e6c3b0a480b442fe57604081701 README.md
2
+ 47be8efd7e1fd31c40bba19c677eed4ced535acdd522f428db29e11aacb39f5d REPRODUCE.md
3
+ 71dc6a4089b81a02634fa40095c900c30173716ec062db4e118fe8a7a66fbec7 data/final81_arms.csv
4
+ 1cf6dea059f599d82be869d5b18db384fc03397630175f3b9202383bb3b09dce data/final81_axes.csv
5
+ c3102a05f489e1a9c16e9a6d1092b857b5a2ac922bc1f4868b42df2fc666bcee data/kernel_verification.json
6
+ 97999690deb70e9619017c834af8849bc93e50bfbc75e6c9fa0d0a0c792fcac4 data/paired_comparisons.csv
7
+ f3513129bf9097315c122a51df2504425fe86b76718f68062da142a6a532ae0c data/run_metadata.json
8
+ be406d1e0b1cdc61a7ac6bc9d3a7af1674b69c8c182a8628c0edb3126da6502f recipes/a81_allocation.json
README.md ADDED
@@ -0,0 +1,132 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: cc-by-4.0
3
+ pretty_name: B200 Low-Bit Quantization Benchmarks (Qwen3-8B, D81)
4
+ language:
5
+ - en
6
+ - ko
7
+ tags:
8
+ - quantization
9
+ - nvfp4
10
+ - mixed-precision
11
+ - blackwell
12
+ - b200
13
+ - vllm
14
+ - benchmark
15
+ task_categories:
16
+ - text-generation
17
+ size_categories:
18
+ - n<1K
19
+ configs:
20
+ - config_name: arms
21
+ data_files: data/final81_arms.csv
22
+ - config_name: axes
23
+ data_files: data/final81_axes.csv
24
+ - config_name: paired
25
+ data_files: data/paired_comparisons.csv
26
+ ---
27
+
28
+ # B200 Low-Bit Quantization Benchmarks (Qwen3-8B)
29
+
30
+ Aggregate results, recipes and serving measurements from the paper
31
+ *Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell*
32
+ (arXiv link: ARXIV_PENDING).
33
+
34
+ This repository holds **results only**. It contains no model weights, no evaluation items, no per-item predictions
35
+ and no calibration text. The quantized checkpoints evaluated in the paper were not retained.
36
+
37
+ A separate build of the A81 recipe is released as weights in
38
+ [ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200](https://huggingface.co/ThakiCloud/Qwen3-8B-NVFP4A16-INT8-Mixed-B200).
39
+ That checkpoint was evaluated on its own, on an independent set (REL-81), and its card reports those results.
40
+ They are not part of the paper's numbers and are not included in the files here.
41
+
42
+ ## What was tested
43
+
44
+ Base model: `Qwen/Qwen3-8B` (Apache-2.0). Hardware: one NVIDIA B200. Serving: vLLM 0.29.0, batch size 1.
45
+ Every quantized configuration had to pass, on a sealed evaluation set it was never tuned on:
46
+
47
+ - runtime weight memory below 5.98 GiB (measured weight-load bytes / 2^30),
48
+ - Korean retention and tier retention of at least 0.90 relative to BF16,
49
+ - overall, numeric and reasoning accuracy no more than 3 points below BF16, and format-error rate no more than
50
+ 2 points above BF16,
51
+ - time per output token at most 1.1 x BF16,
52
+ - every quantized linear layer served by its intended packed kernel (no float fallback).
53
+
54
+ The confirmatory set (FINAL-81) has 5168 items: 1565 Korean, 1043 English, 1320 numeric,
55
+ 870 reasoning, 300 Korean math and 70 Korean in-domain. Its SHA-256 prefix is
56
+ `26a1fadc91b6`. Items are not redistributed.
57
+
58
+ ## Main result
59
+
60
+ | Arm | Memory GiB | Tier ret. | Korean ret. | TTFT p50 ms | TPOT ms | tok/s (b1) | TPOT / BF16 | Gate |
61
+ |---|---|---|---|---|---|---|---|---|
62
+ | BF16 | 15.268 | - | - | 15.251 | 3.9154 | 255.4 | 1.0 | reference |
63
+ | A81 | 5.917 | 0.9311 | 0.9036 | 15.39 | 3.6759 | 272.0 | 0.9388 | PASS |
64
+ | NVFP4 W4A4 | 5.977 | 0.8867 | 0.8208 | 42.26 | 2.5837 | 387.0 | 0.6599 | FAIL |
65
+ | B81 | 5.153 | 0.8103 | 0.6996 | 16.823 | 3.3792 | 295.9 | 0.8631 | FAIL |
66
+ | GPTQ W3 | 4.864 | 0.7477 | 0.6428 | 16.276 | 3.3895 | 295.0 | 0.8657 | FAIL |
67
+
68
+ A81 (NVFP4A16 body, 46 INT8 modules, INT8 `lm_head`) was the only configuration that passed every gate:
69
+ 5.917 GiB (61.2% below BF16), tier retention 0.931, Korean retention 0.904.
70
+ All 144 quantized linear layers ran on the intended kernels (98 MarlinNvFp4, 46 Marlin).
71
+
72
+ ## Paired comparisons
73
+
74
+ Differences are accuracy points on the same items; intervals are simultaneous at the listed z.
75
+
76
+ | Comparison | Set | Diff (pp) | CI | Status |
77
+ |---|---|---|---|---|
78
+ | A81 - NVFP4 W4A4 | FINAL-81 | +1.4 | [+0.2, +2.6] | confirmatory (preregistered) |
79
+ | B81 - GPTQ W3 | FINAL-81 | +4.1 | [+2.3, +5.9] | confirmatory (preregistered) |
80
+ | A81 - NVFP4 W4A4, Korean axis only | FINAL-81 | +1.1 | [-1.1, +3.3] | report-only |
81
+ | A81 - NVFP4 W4A4, reasoning axis only | FINAL-81 | +3.8 | [+0.3, +7.2] | report-only |
82
+ | A81 - NVFP4 W4A4 | SEL-81 | +3.4 | [+2.3, +4.6] | selection set |
83
+ | NVFP4A16 (BF16 head) - NVFP4 W4A4 | DEV-80 | +2.8 | [+1.6, +4.0] | exploratory |
84
+ | NVFP4A16 + INT8 units + INT8 head (D80 17-unit allocation) - NVFP4A16 (BF16 head) | DEV-80 | -0.3 | [-1.3, +0.8] | exploratory |
85
+ | B81 with 50% Korean calibration - B81, Korean axis | SEL-81 | +2.2 | [-0.9, +5.4] | preregistered calibration test |
86
+ | B81 with 50% Korean calibration - B81, reasoning axis | SEL-81 | -12.3 | [-16.8, -7.9] | preregistered calibration test |
87
+
88
+ ## How to read these numbers
89
+
90
+ - **Narrow Korean margin.** A81's Korean retention cleared the 0.90 gate by 0.0036 on 1565 Korean items.
91
+ A different rebuild could land on the other side of the gate.
92
+ - **Attribution is exploratory.** On the development set, moving from NVFP4 W4A4 to 16-bit activations (NVFP4A16)
93
+ accounts for much of the gain (+2.8 points [+1.6, +4.0]). Adding INT8 modules on top changed accuracy by
94
+ -0.3 points [-1.3, +0.8], using an earlier 17-unit allocation. The effect of A81's own 46-unit
95
+ allocation was not identified on an independent set.
96
+ - **Speed trade-off.** NVFP4 W4A4 had the highest batch-1 decode throughput (387 tok/s) but a TTFT of 42.3 ms
97
+ and failed the quality gates. A81 decodes at 272 tok/s with a TTFT of 15.4 ms. These are batch-1 numbers;
98
+ throughput at higher batch sizes was not measured.
99
+ - **Not bitwise reproducible.** The SEL-81 and FINAL-81 builds of A81 had different checkpoint hashes
100
+ (`9c4bce0c5e91` vs `7e6fcc531386`). Quality variance across rebuilds was not measured.
101
+ - **Korean-heavy calibration did not transfer.** Recalibrating B81 with 50% Korean text changed Korean accuracy by
102
+ +2.2 [-0.9, +5.4] (not significant) and reasoning by -12.3 [-16.8, -7.9] on SEL-81. A81 used the default
103
+ calibration mix.
104
+
105
+ ## Files
106
+
107
+ - `data/final81_arms.csv` — memory, quality retention, serving latency and gate verdict per arm.
108
+ - `data/final81_axes.csv` — per-axis aggregate accuracy (BF16 vs quantized) per arm.
109
+ - `data/paired_comparisons.csv` — paired differences with confidence intervals.
110
+ - `data/kernel_verification.json` — kernel assignment per arm, mismatches, float fallbacks.
111
+ - `data/run_metadata.json` — gates, set identity, hardware, serving configuration, build hashes.
112
+ - `recipes/a81_allocation.json` — A81 module-level precision allocation (which projections are INT8).
113
+ - `REPRODUCE.md` — how to rebuild A81 and re-measure it.
114
+ - `MANIFEST.sha256` — file hashes.
115
+
116
+ ## Data sources and license
117
+
118
+ The evaluation sets were assembled from public benchmarks (KMMLU, KoBEST, MMLU, ARC, HellaSwag, BBH, BIG-bench,
119
+ MuSR, HRM8K) and a programmatic numeric generator. This repository republishes none of their items.
120
+ The aggregate results in this repository are released under CC BY 4.0 by ThakiCloud.
121
+ `Qwen3-8B` is distributed by its authors under Apache-2.0.
122
+
123
+ ## Citation
124
+
125
+ ```bibtex
126
+ @misc{han2026beyondbitwidth,
127
+ title = {Beyond Bitwidth: Deployment-Constrained Mixed-Precision Quantization for Multilingual LLM Inference on NVIDIA Blackwell},
128
+ author = {Han, Hyojung},
129
+ year = {2026},
130
+ note = {arXiv: ARXIV_PENDING}
131
+ }
132
+ ```
REPRODUCE.md ADDED
@@ -0,0 +1,30 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Reproducing A81
2
+
3
+ This describes how the A81 configuration was built and measured. It is a recipe, not a guarantee: two builds of
4
+ the same recipe produced different checkpoint hashes, so expect small quality differences between rebuilds.
5
+
6
+ ## Pinned environment
7
+
8
+ - Base model: `Qwen/Qwen3-8B`, revision `b968826d9c46dd6066d109eabc6255188de91218`
9
+ - Quantization: `llm-compressor 0.15.0`, `compressed-tensors 0.20.1` (isolated Python environment)
10
+ - Serving: `vllm/vllm-openai:v0.29.0` image, no batch-invariance flag, no linear-backend override
11
+ - Hardware: 1x NVIDIA B200
12
+ - Chat template with thinking disabled (`enable_thinking=False`)
13
+
14
+ ## Recipe
15
+
16
+ 1. Quantize every decoder linear projection (q, k, v, o, gate, up, down; 36 layers) with GPTQ.
17
+ 2. Default scheme: NVFP4A16 (FP4 E2M1 weights, group size 16, 16-bit activations).
18
+ 3. Override the 46 units listed in `recipes/a81_allocation.json` to INT8 weight-only, symmetric, group size 128.
19
+ A unit `L.qkv` covers `q_proj`, `k_proj`, `v_proj` of layer `L`; `L.gate_up` covers `gate_proj` and `up_proj`.
20
+ 4. Quantize `lm_head` to INT8 weight-only, symmetric, group size 128. Embeddings stay in BF16.
21
+ 5. Calibrate with the default (non-Korean-weighted) calibration mix. The calibration text is not distributed.
22
+
23
+ ## Checks before trusting a build
24
+
25
+ 1. **Kernels.** Load in vLLM and confirm 98 layers on `MarlinNvFp4LinearKernel`, 46 on `MarlinLinearKernel`,
26
+ zero float fallbacks, `lm_head` packed.
27
+ 2. **Memory.** Weight-load bytes / 2^30 must stay below 5.98 GiB (paper build: 5.917 GiB).
28
+ 3. **Latency.** Batch 1, 32 requests: TPOT should be at most 1.1x BF16 on the same GPU.
29
+ 4. **Quality.** Evaluate on an independent set you did not use for any tuning. Do not compare against the paper's
30
+ FINAL-81 numbers as if they were the same build.
data/final81_arms.csv ADDED
@@ -0,0 +1,6 @@
 
 
 
 
 
 
 
1
+ arm,precision,role,runtime_weight_memory_gib,runtime_weight_memory_bytes,memory_reduction_vs_bf16_pct,effective_bpw,tier_retention,korean_retention,accuracy_delta_vs_bf16_pp,ttft_ms_p50,ttft_ms_p95,tpot_ms_median,tokens_per_s_batch1,tpot_ratio_vs_bf16,kernels,kernel_mismatches,product_gate
2
+ BF16,BF16 weights and activations (reference),reference,15.268,16394145792,,,,,,15.251,20.28,3.9154,255.4,1.0,BF16 dense=144,,reference
3
+ A81,"NVFP4A16 body (FP4 E2M1 weights, group 16, 16-bit activations) + 46 INT8 weight-only modules (sym, group 128) + INT8 lm_head",candidate (track A),5.917,6353372672,61.2,6.1732,0.9311,0.9036,-0.31,15.39,16.241,3.6759,272.0,0.9388,MarlinLinearKernel=46;MarlinNvFp4LinearKernel=98,0,PASS
4
+ NVFP4 W4A4,NVFP4 weights and NVFP4 activations (W4A4),fixed baseline for A81,5.977,6417801216,60.9,6.2481,0.8867,0.8208,-1.73,42.26,43.998,2.5837,387.0,0.6599,FlashInferCuteDslNvFp4LinearKernel=144,0,FAIL
5
+ B81,GPTQ INT3 body + 44 INT4 modules (group 128),candidate (track B),5.153,5533509632,66.2,5.3453,0.8103,0.6996,-6.73,16.823,17.536,3.3792,295.9,0.8631,HummingLinearKernel=100;MarlinLinearKernel=44,0,FAIL
6
+ GPTQ W3,"GPTQ INT3 weight-only, group 128",fixed baseline for B81,4.864,5222719488,68.1,5.0821,0.7477,0.6428,-10.86,16.276,16.868,3.3895,295.0,0.8657,HummingLinearKernel=144,0,FAIL
data/final81_axes.csv ADDED
@@ -0,0 +1,25 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ arm,axis,n_items,bf16_accuracy,quantized_accuracy,accuracy_delta_pp,retention
2
+ A81,ko,1565,0.5169,0.5131,-0.38,0.9036
3
+ A81,en,1043,0.6779,0.6817,0.38,0.9901
4
+ A81,number,1320,0.9045,0.8955,-0.91,0.9313
5
+ A81,reasoning,870,0.7391,0.7379,-0.11,0.9005
6
+ A81,ko_math,300,0.47,0.4533,-1.67,0.922
7
+ A81,ko_indomain,70,0.6857,0.7286,4.29,1.0
8
+ NVFP4 W4A4,ko,1565,0.5169,0.5022,-1.47,0.8208
9
+ NVFP4 W4A4,en,1043,0.6779,0.6692,-0.86,0.9547
10
+ NVFP4 W4A4,number,1320,0.9045,0.8917,-1.29,0.9196
11
+ NVFP4 W4A4,reasoning,870,0.7391,0.7,-3.91,0.8336
12
+ NVFP4 W4A4,ko_math,300,0.47,0.4733,0.33,0.9574
13
+ NVFP4 W4A4,ko_indomain,70,0.6857,0.6714,-1.43,0.9583
14
+ B81,ko,1565,0.5169,0.4588,-5.81,0.6996
15
+ B81,en,1043,0.6779,0.6414,-3.64,0.9081
16
+ B81,number,1320,0.9045,0.7826,-12.2,0.8141
17
+ B81,reasoning,870,0.7391,0.7011,-3.79,0.8351
18
+ B81,ko_math,300,0.47,0.4,-7.0,0.7943
19
+ B81,ko_indomain,70,0.6857,0.5714,-11.43,0.8125
20
+ GPTQ W3,ko,1565,0.5169,0.4428,-7.41,0.6428
21
+ GPTQ W3,en,1043,0.6779,0.6213,-5.66,0.8685
22
+ GPTQ W3,number,1320,0.9045,0.697,-20.76,0.7219
23
+ GPTQ W3,reasoning,870,0.7391,0.6563,-8.28,0.7947
24
+ GPTQ W3,ko_math,300,0.47,0.3767,-9.33,0.7518
25
+ GPTQ W3,ko_indomain,70,0.6857,0.5429,-14.29,0.7708
data/kernel_verification.json ADDED
@@ -0,0 +1,40 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "A81": {
3
+ "by_kernel": {
4
+ "MarlinLinearKernel": 46,
5
+ "MarlinNvFp4LinearKernel": 98
6
+ },
7
+ "mismatches": 0,
8
+ "float_fallback_modules": 0,
9
+ "lm_head_ok": true,
10
+ "env_check": "ENV_MATCH"
11
+ },
12
+ "NVFP4 W4A4": {
13
+ "by_kernel": {
14
+ "FlashInferCuteDslNvFp4LinearKernel": 144
15
+ },
16
+ "mismatches": 0,
17
+ "float_fallback_modules": 0,
18
+ "lm_head_ok": null,
19
+ "env_check": "ENV_MATCH"
20
+ },
21
+ "B81": {
22
+ "by_kernel": {
23
+ "HummingLinearKernel": 100,
24
+ "MarlinLinearKernel": 44
25
+ },
26
+ "mismatches": 0,
27
+ "float_fallback_modules": 0,
28
+ "lm_head_ok": true,
29
+ "env_check": "ENV_MATCH"
30
+ },
31
+ "GPTQ W3": {
32
+ "by_kernel": {
33
+ "HummingLinearKernel": 144
34
+ },
35
+ "mismatches": 0,
36
+ "float_fallback_modules": 0,
37
+ "lm_head_ok": true,
38
+ "env_check": "ENV_MATCH"
39
+ }
40
+ }
data/paired_comparisons.csv ADDED
@@ -0,0 +1,10 @@
 
 
 
 
 
 
 
 
 
 
 
1
+ comparison,set,z,status,diff_pp,ci_lo_pp,ci_hi_pp
2
+ A81 - NVFP4 W4A4,FINAL-81,2.2414,confirmatory (preregistered),1.4,0.2,2.6
3
+ B81 - GPTQ W3,FINAL-81,2.5758,confirmatory (preregistered),4.1,2.3,5.9
4
+ "A81 - NVFP4 W4A4, Korean axis only",FINAL-81,2.2414,report-only,1.1,-1.1,3.3
5
+ "A81 - NVFP4 W4A4, reasoning axis only",FINAL-81,2.2414,report-only,3.8,0.3,7.2
6
+ A81 - NVFP4 W4A4,SEL-81,2.2414,selection set,3.4,2.3,4.6
7
+ NVFP4A16 (BF16 head) - NVFP4 W4A4,DEV-80,1.96,exploratory,2.8,1.6,4.0
8
+ NVFP4A16 + INT8 units + INT8 head (D80 17-unit allocation) - NVFP4A16 (BF16 head),DEV-80,1.96,exploratory,-0.3,-1.3,0.8
9
+ "B81 with 50% Korean calibration - B81, Korean axis",SEL-81,2.5758,preregistered calibration test,2.2,-0.9,5.4
10
+ "B81 with 50% Korean calibration - B81, reasoning axis",SEL-81,2.5758,preregistered calibration test,-12.3,-16.8,-7.9
data/run_metadata.json ADDED
@@ -0,0 +1,37 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "eval_set": {
3
+ "name": "FINAL-81",
4
+ "n_items": 5168,
5
+ "n_gated": 4798,
6
+ "sha256": "26a1fadc91b64ffd84d925fbd84fe22faa622a8dc16612615c8f922e0559ea41",
7
+ "axes": {
8
+ "ko": 1565,
9
+ "en": 1043,
10
+ "number": 1320,
11
+ "reasoning": 870,
12
+ "ko_math": 300,
13
+ "ko_indomain": 70
14
+ },
15
+ "note": "items are not distributed; the sha256 identifies the frozen set used in the paper"
16
+ },
17
+ "gates": {
18
+ "ko_ret": 0.9,
19
+ "num_dacc": -0.03,
20
+ "tier": 0.9,
21
+ "dacc": -0.03,
22
+ "reasoning_dacc": -0.03,
23
+ "format_pp": 0.02,
24
+ "mem_gib_below": 5.98,
25
+ "tpot_vs_bf16": 1.1
26
+ },
27
+ "prereg_sha256": "bd5ffe97e4fb02e8c9936bb4f32dfef9c1cb7de97705de254471a169cadea1b3",
28
+ "verdict": "D81_DEPLOY_CANDIDATE",
29
+ "hardware": "1x NVIDIA B200",
30
+ "serving": "vLLM 0.29.0, batch 1, 32 requests",
31
+ "checkpoint": {
32
+ "final81_build_sha256_prefix": "7e6fcc531386",
33
+ "sel81_build_sha256_prefix": "9c4bce0c5e91",
34
+ "final81_checkpoint_bytes": 6320384652,
35
+ "note": "rebuilds were not bitwise identical; neither checkpoint was retained"
36
+ }
37
+ }
recipes/a81_allocation.json ADDED
@@ -0,0 +1,80 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "arm": "A81",
3
+ "base_model": "Qwen/Qwen3-8B",
4
+ "base_revision": "b968826d9c46dd6066d109eabc6255188de91218",
5
+ "algorithm": "GPTQ (llm-compressor 0.15.0, compressed-tensors 0.20.1)",
6
+ "body": {
7
+ "format": "NVFP4A16",
8
+ "weights": "FP4 E2M1, group 16, FP8 scale",
9
+ "activations": "16-bit"
10
+ },
11
+ "int8_units": {
12
+ "format": "INT8 weight-only, symmetric, group 128",
13
+ "count": 46,
14
+ "by_type": {
15
+ "qkv": 23,
16
+ "o": 18,
17
+ "gate_up": 1,
18
+ "down": 4
19
+ },
20
+ "unit_naming": "<layer>.<proj>; qkv = q_proj+k_proj+v_proj, gate_up = gate_proj+up_proj",
21
+ "units": [
22
+ "0.qkv",
23
+ "13.qkv",
24
+ "5.qkv",
25
+ "14.qkv",
26
+ "10.o",
27
+ "6.o",
28
+ "14.o",
29
+ "7.o",
30
+ "2.o",
31
+ "12.o",
32
+ "11.qkv",
33
+ "17.qkv",
34
+ "18.qkv",
35
+ "8.qkv",
36
+ "2.qkv",
37
+ "1.qkv",
38
+ "4.qkv",
39
+ "0.down",
40
+ "12.qkv",
41
+ "11.down",
42
+ "20.o",
43
+ "15.o",
44
+ "4.o",
45
+ "21.o",
46
+ "17.o",
47
+ "5.o",
48
+ "1.o",
49
+ "9.o",
50
+ "8.o",
51
+ "13.o",
52
+ "11.o",
53
+ "19.o",
54
+ "5.down",
55
+ "24.qkv",
56
+ "9.gate_up",
57
+ "16.qkv",
58
+ "21.qkv",
59
+ "19.qkv",
60
+ "10.qkv",
61
+ "7.qkv",
62
+ "20.qkv",
63
+ "15.qkv",
64
+ "6.qkv",
65
+ "3.qkv",
66
+ "9.qkv",
67
+ "8.down"
68
+ ]
69
+ },
70
+ "lm_head": "INT8 weight-only, symmetric, group 128",
71
+ "memory_budget_gib": 0.50111,
72
+ "units_gib": 0.49915,
73
+ "predicted_runtime_gib": 5.94678,
74
+ "limit_gib": 5.94875,
75
+ "selection": "units chosen on SEL-81 by joint (Korean + reasoning) sensitivity rank under the memory budget; frozen before FINAL-81",
76
+ "kernels_expected": {
77
+ "MarlinNvFp4LinearKernel": 98,
78
+ "MarlinLinearKernel": 46
79
+ }
80
+ }