Hugging Face's logo

endnai
/
DeepSeek-V4-Flash-W4A8-FP8

Text Generation
Transformers
Safetensors
deepseek_v4
Mixture of Experts
compressed-tensors
w4a8
int4
fp8
vllm
quantization
Model card Files
xet
 Community
endnai commited on
Commit
ca92d50
·
verified ·
1 Parent(s): 53cfe7c

W4A8 config/recipe + investigation model card (weights identical to base)

Browse files
Files changed (6) hide show
  1. README.md +135 -0
  2. config.json +1177 -0
  3. generation_config.json +9 -0
  4. model.safetensors.index.json +0 -0
  5. tokenizer.json +0 -0
  6. tokenizer_config.json +14 -0
README.md ADDED
@@ -0,0 +1,135 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: mit
3
+ base_model:
4
+ - canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP
5
+ - deepseek-ai/DeepSeek-V4-Flash
6
+ tags:
7
+ - deepseek_v4
8
+ - moe
9
+ - compressed-tensors
10
+ - w4a8
11
+ - int4
12
+ - fp8
13
+ - vllm
14
+ - quantization
15
+ library_name: transformers
16
+ pipeline_tag: text-generation
17
+ ---
18
+
19
+ # DeepSeek-V4-Flash — W4A8 (INT4 weights + FP8 dynamic-token activations)
20
+
21
+ A **W4A8** quantization of DeepSeek-V4-Flash: **INT4 group-quantized MoE expert weights** with **FP8 (e4m3) dynamic per-token activations**, plus FP8 block-quantized attention/dense layers. Produced as a **zero-cost config transformation** of [`canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP`](https://huggingface.co/canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP) — the INT4 weight bytes are **identical**; only the activation quantization scheme in `config.json` changed (experts `input_activations`: `null` → FP8 dynamic-token).
22
+
23
+ > **⚠️ Honest headline first:** on H200 (Hopper / SM90) this checkpoint **does not make prefill or decode faster than the W4A16 base it was derived from.** It serves correctly and is footprint-neutral (same INT4 weights, same TP2), but W4A8 ≈ W4A16 in throughput. It is published as a **reproducible research artifact** documenting *why* the activation-precision lever doesn't move DeepSeek-V4-Flash performance on Hopper. See **[Investigation & findings](#investigation--findings)**.
24
+
25
+ > **📦 This is a config / recipe repository — the weight shards are NOT included.** Because the W4A8 transformation reuses the base's INT4 weights **byte-for-byte**, duplicating ~159 GB here would be pure waste. This repo ships the W4A8 `config.json`, tokenizer, weight index, and this card. To get a runnable checkpoint, pull the weights from the base and drop in this `config.json` — see **[Getting the weights](#getting-the-weights)** (one command).
26
+
27
+ ## What this is
28
+
29
+ | | |
30
+ |---|---|
31
+ | Base architecture | DeepSeek-V4-Flash (284B total / ~13B active MoE, 43 layers, 256 routed experts top-6 + 1 shared, MLA, hybrid sparse attention + Lightning indexer) |
32
+ | Derived from | `canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP` (identical INT4 expert weights) |
33
+ | MoE experts | **INT4** group-quantized weights + **FP8 e4m3 dynamic per-token** activations (W4A8) |
34
+ | Attention / dense | FP8 block-quantized weights (unchanged from base) |
35
+ | `format` | `mixed-precision` (compressed-tensors) |
36
+ | Footprint | ~159 GB materialized, fits **TP2** on 2×H200 (identical to the W4A16 base). **Weights not stored here** — see [Getting the weights](#getting-the-weights). |
37
+ | Target hardware | NVIDIA Hopper (H100/H200, SM90) |
38
+
39
+ ## How it was made
40
+
41
+ DeepSeek-V4-Flash's MoE experts are stored as INT4. A W4A16 checkpoint runs those INT4 weights through a Marlin dequant→BF16 GEMM; a **W4A8** checkpoint instead pairs the *same* INT4 weights with FP8 activations, so vLLM dispatches them to the native **`CutlassExpertsW4A8Fp8`** kernel on SM90 (`_is_fp8_w4a8_sm90`).
42
+
43
+ Because the weights are unchanged, the conversion is a **pure `config.json` edit** — no re-quantization, no calibration:
44
+
45
+ ```jsonc
46
+ // experts config group, input_activations: null ->
47
+ "input_activations": {
48
+ "num_bits": 8, "type": "float", "strategy": "token",
49
+ "dynamic": true, "symmetric": true
50
+ }
51
+ ```
52
+
53
+ The `_w4a8_conversion` key in `config.json` records this provenance.
54
+
55
+ ## Getting the weights
56
+
57
+ The INT4 weight shards are identical to the base. Materialize a full checkpoint by downloading the base weights and overwriting `config.json` with this repo's W4A8 config:
58
+
59
+ ```bash
60
+ # 1. base weights (INT4 shards, tokenizer) — the actual ~159 GB
61
+ hf download canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP --local-dir dsv4-w4a8
62
+
63
+ # 2. this repo's W4A8 config + card (the only real diff)
64
+ hf download endnai/DeepSeek-V4-Flash-W4A8-FP8 config.json README.md --local-dir dsv4-w4a8
65
+
66
+ # dsv4-w4a8/ is now a complete W4A8 checkpoint (INT4 weights + FP8-activation config)
67
+ ```
68
+
69
+ The `.safetensors` bytes are unchanged; only `config.json`'s expert `input_activations` differ (see below).
70
+
71
+ ## Serving (vLLM)
72
+
73
+ Requires a recent **vLLM nightly** and, at the time of writing, four small patches to load the DeepSeek-V4-Flash compressed-tensors checkpoint (these are model-loading fixes, not W4A8-specific — the same patches are needed for the W4A16 base on nightly):
74
+
75
+ 1. `packed_modules_mapping` for the model **and** MTP module (`fused_wqa_wkv`, `fused_wkv_wgate`, `gate_up_proj`).
76
+ 2. `hash_moe` added to the transformers `ALLOWED_LAYER_TYPES` global allowlist.
77
+ 3. `o_proj` weight-scale name alias (`weight_scale_inv` → `weight_scale`).
78
+
79
+ Launch (2×H200, TP2) from the materialized directory (see [Getting the weights](#getting-the-weights)):
80
+
81
+ ```bash
82
+ vllm serve ./dsv4-w4a8 \
83
+ --tensor-parallel-size 2 \
84
+ --disable-custom-all-reduce \
85
+ --trust-remote-code
86
+ ```
87
+
88
+ `--disable-custom-all-reduce` avoids a TP2 init hang under confidential-compute (custom all-reduce needs CUDA-IPC/symmetric memory, which is unavailable inside TDX CVMs).
89
+
90
+ **Correctness:** verified matching the W4A16 base on a temp=0 quality probe (GSM8K 3/3 identical).
91
+
92
+ ## Investigation & findings
93
+
94
+ This checkpoint was built to test a hypothesis: *the DeepSeek-V4-Flash prefill bottleneck is the INT4→BF16 Marlin MoE GEMM, so a W4A8 path (native FP8 activation GEMM) should be ~1.5–2× faster.* **The hypothesis was refuted.** Full sweep on 2–8×H200 (TP2 unless noted), single-request prefill ladder (c=1), long-context (ISL up to 24k):
95
+
96
+ ### Headline: W4A8 gives no throughput advantage over W4A16
97
+
98
+ | Config | Engine | TP | Prefill TTFT @24k | Prefill tok/s/GPU @24k |
99
+ |---|---|---|---|---|
100
+ | **W4A8** (this model) | vLLM | 2 | **1658 ms** | **7410** |
101
+ | W4A16 (base) | vLLM | 2 | 1691 ms | 7267 |
102
+ | FP4 (marlin) | vLLM | 2 | 1824 ms | 7090 |
103
+ | FP4 (marlin) | sglang | 2 | 1894 ms | 6832 |
104
+ | FP8 (native) | sglang | 4 | 892 ms | 6888 |
105
+
106
+ **Per-GPU prefill throughput is flat at ~6.8–7.4k tok/s/GPU across every engine and every quantization.** W4A8 and W4A16 are a **wash** (1658 vs 1691 ms — within noise). The FP8-TP4 config's lower absolute TTFT (892 ms) is **pure tensor-parallel scaling** (2× the GPUs); per-GPU it is *also* a wash.
107
+
108
+ ### Why the activation-precision lever doesn't help
109
+
110
+ At prefill batch sizes, the DeepSeek-V4-Flash MoE (top-6 of 256 small experts) is **weight-bandwidth-bound**, not compute-bound on the expert GEMM. INT4 weights are already the bandwidth-optimal format, and Marlin's INT4→BF16 path already matches the Cutlass W4A8 kernel in practice. Switching activations from BF16/FP8-implicit to FP8 changes the *activation* precision but not the dominant cost. The compute-bound portion of prefill is dominated by **format-shared work** — FP8-block MLA attention and the sparse / Lightning-indexer passes over long context — which is identical across all three checkpoints.
111
+
112
+ ### The prefill ceiling is architectural on Hopper
113
+
114
+ - Prefill scales **linearly** above ~8k tokens (~+547 ms per +8k) with GPUs at ~100% util and ~690 W (near TDP) → tensor-core-bound, not launch- or attention-quadratic-bound.
115
+ - The two kernel improvements that *would* help — **native NVFP4 MoE GEMM** and the **FP4 Lightning-indexer cache** — are **Blackwell-only** (SM100). On Hopper, sglang/vLLM fall back to Marlin.
116
+ - A **W4A8 SM90 grouped-GEMM** tuned for the DeepSeek-V4 MoE path is unimplemented upstream (relevant issues closed inactive). Even so, the wash above suggests it would offer little at prefill batch-M.
117
+
118
+ ### What *does* move the needle (deployment)
119
+
120
+ - **Prefix caching** is the dominant lever: in production, DeepSeek-V4-Flash realizes **~55% radix prefix-cache hit** on real agent/RAG traffic (measured over 24h), i.e. more than half of all prefill is skipped. This is already captured by sglang RadixAttention in production.
121
+ - Larger **chunked-prefill** (8192 → 16384) gives ~7% faster long-context prefill TTFT on sglang, at the cost of KV-concurrency — a free win when the server isn't KV-bound.
122
+
123
+ ### Bottom line
124
+
125
+ Use W4A8 for **GPU-footprint efficiency** (TP2, ~159 GB, native FP8 activation compute where a downstream kernel benefits) — but **not** expecting it to beat W4A16 on DeepSeek-V4-Flash prefill/decode on Hopper. On this architecture the two are equivalent; the real gains come from prefix caching and, eventually, Blackwell hardware.
126
+
127
+ ## Reproducibility
128
+
129
+ - Weights: byte-identical to `canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP`.
130
+ - Transformation: the single `config.json` `input_activations` edit shown above (see the `_w4a8_conversion` provenance key).
131
+ - To rebuild: take the W4A16 base, apply the config edit, serve with the vLLM nightly + patches above.
132
+
133
+ ## Acknowledgements
134
+
135
+ Built and benchmarked by Evrard Nil with Claude (2026-06). Base quantization by `canada-quant`; original model by DeepSeek-AI.
config.json ADDED
@@ -0,0 +1,1177 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "architectures": [
3
+ "DeepseekV4ForCausalLM"
4
+ ],
5
+ "attention_bias": false,
6
+ "attention_dropout": 0.0,
7
+ "bos_token_id": 0,
8
+ "compress_rates": {
9
+ "compressed_sparse_attention": 4,
10
+ "heavily_compressed_attention": 128
11
+ },
12
+ "compress_rope_theta": 160000,
13
+ "dtype": "bfloat16",
14
+ "eos_token_id": 1,
15
+ "expert_dtype": "bf16",
16
+ "hc_eps": 1e-06,
17
+ "hc_mult": 4,
18
+ "hc_sinkhorn_iters": 20,
19
+ "head_dim": 512,
20
+ "hidden_act": "silu",
21
+ "hidden_size": 4096,
22
+ "index_head_dim": 128,
23
+ "index_n_heads": 64,
24
+ "index_topk": 512,
25
+ "initializer_range": 0.02,
26
+ "layer_types": [
27
+ "sliding_attention",
28
+ "sliding_attention",
29
+ "compressed_sparse_attention",
30
+ "heavily_compressed_attention",
31
+ "compressed_sparse_attention",
32
+ "heavily_compressed_attention",
33
+ "compressed_sparse_attention",
34
+ "heavily_compressed_attention",
35
+ "compressed_sparse_attention",
36
+ "heavily_compressed_attention",
37
+ "compressed_sparse_attention",
38
+ "heavily_compressed_attention",
39
+ "compressed_sparse_attention",
40
+ "heavily_compressed_attention",
41
+ "compressed_sparse_attention",
42
+ "heavily_compressed_attention",
43
+ "compressed_sparse_attention",
44
+ "heavily_compressed_attention",
45
+ "compressed_sparse_attention",
46
+ "heavily_compressed_attention",
47
+ "compressed_sparse_attention",
48
+ "heavily_compressed_attention",
49
+ "compressed_sparse_attention",
50
+ "heavily_compressed_attention",
51
+ "compressed_sparse_attention",
52
+ "heavily_compressed_attention",
53
+ "compressed_sparse_attention",
54
+ "heavily_compressed_attention",
55
+ "compressed_sparse_attention",
56
+ "heavily_compressed_attention",
57
+ "compressed_sparse_attention",
58
+ "heavily_compressed_attention",
59
+ "compressed_sparse_attention",
60
+ "heavily_compressed_attention",
61
+ "compressed_sparse_attention",
62
+ "heavily_compressed_attention",
63
+ "compressed_sparse_attention",
64
+ "heavily_compressed_attention",
65
+ "compressed_sparse_attention",
66
+ "heavily_compressed_attention",
67
+ "compressed_sparse_attention",
68
+ "heavily_compressed_attention",
69
+ "compressed_sparse_attention"
70
+ ],
71
+ "max_position_embeddings": 1048576,
72
+ "mlp_bias": false,
73
+ "mlp_layer_types": [
74
+ "hash_moe",
75
+ "hash_moe",
76
+ "hash_moe",
77
+ "moe",
78
+ "moe",
79
+ "moe",
80
+ "moe",
81
+ "moe",
82
+ "moe",
83
+ "moe",
84
+ "moe",
85
+ "moe",
86
+ "moe",
87
+ "moe",
88
+ "moe",
89
+ "moe",
90
+ "moe",
91
+ "moe",
92
+ "moe",
93
+ "moe",
94
+ "moe",
95
+ "moe",
96
+ "moe",
97
+ "moe",
98
+ "moe",
99
+ "moe",
100
+ "moe",
101
+ "moe",
102
+ "moe",
103
+ "moe",
104
+ "moe",
105
+ "moe",
106
+ "moe",
107
+ "moe",
108
+ "moe",
109
+ "moe",
110
+ "moe",
111
+ "moe",
112
+ "moe",
113
+ "moe",
114
+ "moe",
115
+ "moe",
116
+ "moe"
117
+ ],
118
+ "model_type": "deepseek_v4",
119
+ "moe_intermediate_size": 2048,
120
+ "n_routed_experts": 256,
121
+ "n_shared_experts": 1,
122
+ "norm_topk_prob": true,
123
+ "num_attention_heads": 64,
124
+ "num_experts_per_tok": 6,
125
+ "num_hidden_layers": 43,
126
+ "num_key_value_heads": 1,
127
+ "num_nextn_predict_layers": 1,
128
+ "o_groups": 8,
129
+ "o_lora_rank": 1024,
130
+ "output_router_logits": false,
131
+ "pad_token_id": null,
132
+ "partial_rotary_factor": 0.125,
133
+ "q_lora_rank": 1024,
134
+ "qk_rope_head_dim": 64,
135
+ "quantization_config": {
136
+ "config_groups": {
137
+ "group_0": {
138
+ "format": "float-quantized",
139
+ "input_activations": {
140
+ "actorder": null,
141
+ "block_structure": null,
142
+ "dynamic": true,
143
+ "group_size": 128,
144
+ "num_bits": 8,
145
+ "observer": null,
146
+ "observer_kwargs": {},
147
+ "scale_dtype": null,
148
+ "strategy": "group",
149
+ "symmetric": true,
150
+ "type": "float",
151
+ "zp_dtype": null
152
+ },
153
+ "output_activations": null,
154
+ "targets": [
155
+ "re:.*attn\\.(wq_a|wq_b|wkv|wo_a|wo_b|fused_wqa_wkv|q_a_proj|q_b_proj|kv_proj|o_a_proj|o_b_proj)$"
156
+ ],
157
+ "weights": {
158
+ "actorder": null,
159
+ "block_structure": [
160
+ 128,
161
+ 128
162
+ ],
163
+ "dynamic": false,
164
+ "group_size": null,
165
+ "num_bits": 8,
166
+ "observer": "memoryless_minmax",
167
+ "observer_kwargs": {},
168
+ "scale_dtype": null,
169
+ "strategy": "block",
170
+ "symmetric": true,
171
+ "type": "float",
172
+ "zp_dtype": null
173
+ }
174
+ },
175
+ "group_1": {
176
+ "format": "pack-quantized",
177
+ "input_activations": {
178
+ "num_bits": 8,
179
+ "type": "float",
180
+ "strategy": "token",
181
+ "dynamic": true,
182
+ "symmetric": true,
183
+ "group_size": null,
184
+ "block_structure": null,
185
+ "actorder": null,
186
+ "observer": null,
187
+ "observer_kwargs": {},
188
+ "scale_dtype": null,
189
+ "zp_dtype": null
190
+ },
191
+ "output_activations": null,
192
+ "targets": [
193
+ "re:.*experts\\.\\d+\\.(w1|w2|w3|gate_proj|up_proj|down_proj|gate_up_proj)$"
194
+ ],
195
+ "weights": {
196
+ "actorder": "static",
197
+ "block_structure": null,
198
+ "dynamic": false,
199
+ "group_size": 128,
200
+ "num_bits": 4,
201
+ "observer": "memoryless_minmax",
202
+ "observer_kwargs": {},
203
+ "scale_dtype": null,
204
+ "strategy": "group",
205
+ "symmetric": true,
206
+ "type": "int",
207
+ "zp_dtype": null
208
+ }
209
+ }
210
+ },
211
+ "format": "mixed-precision",
212
+ "global_compression_ratio": null,
213
+ "ignore": [
214
+ "layers.0.ffn.shared_experts.w1",
215
+ "layers.0.ffn.shared_experts.w2",
216
+ "layers.0.ffn.shared_experts.w3",
217
+ "layers.1.ffn.shared_experts.w1",
218
+ "layers.1.ffn.shared_experts.w2",
219
+ "layers.1.ffn.shared_experts.w3",
220
+ "layers.2.ffn.shared_experts.w1",
221
+ "layers.2.ffn.shared_experts.w2",
222
+ "layers.2.ffn.shared_experts.w3",
223
+ "layers.3.ffn.shared_experts.w1",
224
+ "layers.3.ffn.shared_experts.w2",
225
+ "layers.3.ffn.shared_experts.w3",
226
+ "layers.4.ffn.shared_experts.w1",
227
+ "layers.4.ffn.shared_experts.w2",
228
+ "layers.4.ffn.shared_experts.w3",
229
+ "layers.5.ffn.shared_experts.w1",
230
+ "layers.5.ffn.shared_experts.w2",
231
+ "layers.5.ffn.shared_experts.w3",
232
+ "layers.6.ffn.shared_experts.w1",
233
+ "layers.6.ffn.shared_experts.w2",
234
+ "layers.6.ffn.shared_experts.w3",
235
+ "layers.7.ffn.shared_experts.w1",
236
+ "layers.7.ffn.shared_experts.w2",
237
+ "layers.7.ffn.shared_experts.w3",
238
+ "layers.8.ffn.shared_experts.w1",
239
+ "layers.8.ffn.shared_experts.w2",
240
+ "layers.8.ffn.shared_experts.w3",
241
+ "layers.9.ffn.shared_experts.w1",
242
+ "layers.9.ffn.shared_experts.w2",
243
+ "layers.9.ffn.shared_experts.w3",
244
+ "layers.10.ffn.shared_experts.w1",
245
+ "layers.10.ffn.shared_experts.w2",
246
+ "layers.10.ffn.shared_experts.w3",
247
+ "layers.11.ffn.shared_experts.w1",
248
+ "layers.11.ffn.shared_experts.w2",
249
+ "layers.11.ffn.shared_experts.w3",
250
+ "layers.12.ffn.shared_experts.w1",
251
+ "layers.12.ffn.shared_experts.w2",
252
+ "layers.12.ffn.shared_experts.w3",
253
+ "layers.13.ffn.shared_experts.w1",
254
+ "layers.13.ffn.shared_experts.w2",
255
+ "layers.13.ffn.shared_experts.w3",
256
+ "layers.14.ffn.shared_experts.w1",
257
+ "layers.14.ffn.shared_experts.w2",
258
+ "layers.14.ffn.shared_experts.w3",
259
+ "layers.15.ffn.shared_experts.w1",
260
+ "layers.15.ffn.shared_experts.w2",
261
+ "layers.15.ffn.shared_experts.w3",
262
+ "layers.16.ffn.shared_experts.w1",
263
+ "layers.16.ffn.shared_experts.w2",
264
+ "layers.16.ffn.shared_experts.w3",
265
+ "layers.17.ffn.shared_experts.w1",
266
+ "layers.17.ffn.shared_experts.w2",
267
+ "layers.17.ffn.shared_experts.w3",
268
+ "layers.18.ffn.shared_experts.w1",
269
+ "layers.18.ffn.shared_experts.w2",
270
+ "layers.18.ffn.shared_experts.w3",
271
+ "layers.19.ffn.shared_experts.w1",
272
+ "layers.19.ffn.shared_experts.w2",
273
+ "layers.19.ffn.shared_experts.w3",
274
+ "layers.20.ffn.shared_experts.w1",
275
+ "layers.20.ffn.shared_experts.w2",
276
+ "layers.20.ffn.shared_experts.w3",
277
+ "layers.21.ffn.shared_experts.w1",
278
+ "layers.21.ffn.shared_experts.w2",
279
+ "layers.21.ffn.shared_experts.w3",
280
+ "layers.22.ffn.shared_experts.w1",
281
+ "layers.22.ffn.shared_experts.w2",
282
+ "layers.22.ffn.shared_experts.w3",
283
+ "layers.23.ffn.shared_experts.w1",
284
+ "layers.23.ffn.shared_experts.w2",
285
+ "layers.23.ffn.shared_experts.w3",
286
+ "layers.24.ffn.shared_experts.w1",
287
+ "layers.24.ffn.shared_experts.w2",
288
+ "layers.24.ffn.shared_experts.w3",
289
+ "layers.25.ffn.shared_experts.w1",
290
+ "layers.25.ffn.shared_experts.w2",
291
+ "layers.25.ffn.shared_experts.w3",
292
+ "layers.26.ffn.shared_experts.w1",
293
+ "layers.26.ffn.shared_experts.w2",
294
+ "layers.26.ffn.shared_experts.w3",
295
+ "layers.27.ffn.shared_experts.w1",
296
+ "layers.27.ffn.shared_experts.w2",
297
+ "layers.27.ffn.shared_experts.w3",
298
+ "layers.28.ffn.shared_experts.w1",
299
+ "layers.28.ffn.shared_experts.w2",
300
+ "layers.28.ffn.shared_experts.w3",
301
+ "layers.29.ffn.shared_experts.w1",
302
+ "layers.29.ffn.shared_experts.w2",
303
+ "layers.29.ffn.shared_experts.w3",
304
+ "layers.30.ffn.shared_experts.w1",
305
+ "layers.30.ffn.shared_experts.w2",
306
+ "layers.30.ffn.shared_experts.w3",
307
+ "layers.31.ffn.shared_experts.w1",
308
+ "layers.31.ffn.shared_experts.w2",
309
+ "layers.31.ffn.shared_experts.w3",
310
+ "layers.32.ffn.shared_experts.w1",
311
+ "layers.32.ffn.shared_experts.w2",
312
+ "layers.32.ffn.shared_experts.w3",
313
+ "layers.33.ffn.shared_experts.w1",
314
+ "layers.33.ffn.shared_experts.w2",
315
+ "layers.33.ffn.shared_experts.w3",
316
+ "layers.34.ffn.shared_experts.w1",
317
+ "layers.34.ffn.shared_experts.w2",
318
+ "layers.34.ffn.shared_experts.w3",
319
+ "layers.35.ffn.shared_experts.w1",
320
+ "layers.35.ffn.shared_experts.w2",
321
+ "layers.35.ffn.shared_experts.w3",
322
+ "layers.36.ffn.shared_experts.w1",
323
+ "layers.36.ffn.shared_experts.w2",
324
+ "layers.36.ffn.shared_experts.w3",
325
+ "layers.37.ffn.shared_experts.w1",
326
+ "layers.37.ffn.shared_experts.w2",
327
+ "layers.37.ffn.shared_experts.w3",
328
+ "layers.38.ffn.shared_experts.w1",
329
+ "layers.38.ffn.shared_experts.w2",
330
+ "layers.38.ffn.shared_experts.w3",
331
+ "layers.39.ffn.shared_experts.w1",
332
+ "layers.39.ffn.shared_experts.w2",
333
+ "layers.39.ffn.shared_experts.w3",
334
+ "layers.40.ffn.shared_experts.w1",
335
+ "layers.40.ffn.shared_experts.w2",
336
+ "layers.40.ffn.shared_experts.w3",
337
+ "layers.41.ffn.shared_experts.w1",
338
+ "layers.41.ffn.shared_experts.w2",
339
+ "layers.41.ffn.shared_experts.w3",
340
+ "layers.42.ffn.shared_experts.w1",
341
+ "layers.42.ffn.shared_experts.w2",
342
+ "layers.42.ffn.shared_experts.w3",
343
+ "re:^layers\\.43\\.",
344
+ "re:^model\\.layers\\.43\\.",
345
+ "lm_head",
346
+ "layers.0.attn.compressor.wgate",
347
+ "layers.0.attn.compressor.wkv",
348
+ "layers.0.attn.compressor.fused_wkv_wgate",
349
+ "layers.0.attn.compressor.gate_proj",
350
+ "layers.0.attn.compressor.kv_proj",
351
+ "layers.0.attn.indexer.weights_proj",
352
+ "layers.0.attn.indexer.wq_b",
353
+ "layers.0.attn.indexer.q_b_proj",
354
+ "layers.0.attn.indexer.compressor.wgate",
355
+ "layers.0.attn.indexer.compressor.wkv",
356
+ "layers.0.attn.indexer.compressor.gate_proj",
357
+ "layers.0.attn.indexer.compressor.kv_proj",
358
+ "layers.1.attn.compressor.wgate",
359
+ "layers.1.attn.compressor.wkv",
360
+ "layers.1.attn.compressor.fused_wkv_wgate",
361
+ "layers.1.attn.compressor.gate_proj",
362
+ "layers.1.attn.compressor.kv_proj",
363
+ "layers.1.attn.indexer.weights_proj",
364
+ "layers.1.attn.indexer.wq_b",
365
+ "layers.1.attn.indexer.q_b_proj",
366
+ "layers.1.attn.indexer.compressor.wgate",
367
+ "layers.1.attn.indexer.compressor.wkv",
368
+ "layers.1.attn.indexer.compressor.gate_proj",
369
+ "layers.1.attn.indexer.compressor.kv_proj",
370
+ "layers.2.attn.compressor.wgate",
371
+ "layers.2.attn.compressor.wkv",
372
+ "layers.2.attn.compressor.fused_wkv_wgate",
373
+ "layers.2.attn.compressor.gate_proj",
374
+ "layers.2.attn.compressor.kv_proj",
375
+ "layers.2.attn.indexer.weights_proj",
376
+ "layers.2.attn.indexer.wq_b",
377
+ "layers.2.attn.indexer.q_b_proj",
378
+ "layers.2.attn.indexer.compressor.wgate",
379
+ "layers.2.attn.indexer.compressor.wkv",
380
+ "layers.2.attn.indexer.compressor.gate_proj",
381
+ "layers.2.attn.indexer.compressor.kv_proj",
382
+ "layers.3.attn.compressor.wgate",
383
+ "layers.3.attn.compressor.wkv",
384
+ "layers.3.attn.compressor.fused_wkv_wgate",
385
+ "layers.3.attn.compressor.gate_proj",
386
+ "layers.3.attn.compressor.kv_proj",
387
+ "layers.3.attn.indexer.weights_proj",
388
+ "layers.3.attn.indexer.wq_b",
389
+ "layers.3.attn.indexer.q_b_proj",
390
+ "layers.3.attn.indexer.compressor.wgate",
391
+ "layers.3.attn.indexer.compressor.wkv",
392
+ "layers.3.attn.indexer.compressor.gate_proj",
393
+ "layers.3.attn.indexer.compressor.kv_proj",
394
+ "layers.4.attn.compressor.wgate",
395
+ "layers.4.attn.compressor.wkv",
396
+ "layers.4.attn.compressor.fused_wkv_wgate",
397
+ "layers.4.attn.compressor.gate_proj",
398
+ "layers.4.attn.compressor.kv_proj",
399
+ "layers.4.attn.indexer.weights_proj",
400
+ "layers.4.attn.indexer.wq_b",
401
+ "layers.4.attn.indexer.q_b_proj",
402
+ "layers.4.attn.indexer.compressor.wgate",
403
+ "layers.4.attn.indexer.compressor.wkv",
404
+ "layers.4.attn.indexer.compressor.gate_proj",
405
+ "layers.4.attn.indexer.compressor.kv_proj",
406
+ "layers.5.attn.compressor.wgate",
407
+ "layers.5.attn.compressor.wkv",
408
+ "layers.5.attn.compressor.fused_wkv_wgate",
409
+ "layers.5.attn.compressor.gate_proj",
410
+ "layers.5.attn.compressor.kv_proj",
411
+ "layers.5.attn.indexer.weights_proj",
412
+ "layers.5.attn.indexer.wq_b",
413
+ "layers.5.attn.indexer.q_b_proj",
414
+ "layers.5.attn.indexer.compressor.wgate",
415
+ "layers.5.attn.indexer.compressor.wkv",
416
+ "layers.5.attn.indexer.compressor.gate_proj",
417
+ "layers.5.attn.indexer.compressor.kv_proj",
418
+ "layers.6.attn.compressor.wgate",
419
+ "layers.6.attn.compressor.wkv",
420
+ "layers.6.attn.compressor.fused_wkv_wgate",
421
+ "layers.6.attn.compressor.gate_proj",
422
+ "layers.6.attn.compressor.kv_proj",
423
+ "layers.6.attn.indexer.weights_proj",
424
+ "layers.6.attn.indexer.wq_b",
425
+ "layers.6.attn.indexer.q_b_proj",
426
+ "layers.6.attn.indexer.compressor.wgate",
427
+ "layers.6.attn.indexer.compressor.wkv",
428
+ "layers.6.attn.indexer.compressor.gate_proj",
429
+ "layers.6.attn.indexer.compressor.kv_proj",
430
+ "layers.7.attn.compressor.wgate",
431
+ "layers.7.attn.compressor.wkv",
432
+ "layers.7.attn.compressor.fused_wkv_wgate",
433
+ "layers.7.attn.compressor.gate_proj",
434
+ "layers.7.attn.compressor.kv_proj",
435
+ "layers.7.attn.indexer.weights_proj",
436
+ "layers.7.attn.indexer.wq_b",
437
+ "layers.7.attn.indexer.q_b_proj",
438
+ "layers.7.attn.indexer.compressor.wgate",
439
+ "layers.7.attn.indexer.compressor.wkv",
440
+ "layers.7.attn.indexer.compressor.gate_proj",
441
+ "layers.7.attn.indexer.compressor.kv_proj",
442
+ "layers.8.attn.compressor.wgate",
443
+ "layers.8.attn.compressor.wkv",
444
+ "layers.8.attn.compressor.fused_wkv_wgate",
445
+ "layers.8.attn.compressor.gate_proj",
446
+ "layers.8.attn.compressor.kv_proj",
447
+ "layers.8.attn.indexer.weights_proj",
448
+ "layers.8.attn.indexer.wq_b",
449
+ "layers.8.attn.indexer.q_b_proj",
450
+ "layers.8.attn.indexer.compressor.wgate",
451
+ "layers.8.attn.indexer.compressor.wkv",
452
+ "layers.8.attn.indexer.compressor.gate_proj",
453
+ "layers.8.attn.indexer.compressor.kv_proj",
454
+ "layers.9.attn.compressor.wgate",
455
+ "layers.9.attn.compressor.wkv",
456
+ "layers.9.attn.compressor.fused_wkv_wgate",
457
+ "layers.9.attn.compressor.gate_proj",
458
+ "layers.9.attn.compressor.kv_proj",
459
+ "layers.9.attn.indexer.weights_proj",
460
+ "layers.9.attn.indexer.wq_b",
461
+ "layers.9.attn.indexer.q_b_proj",
462
+ "layers.9.attn.indexer.compressor.wgate",
463
+ "layers.9.attn.indexer.compressor.wkv",
464
+ "layers.9.attn.indexer.compressor.gate_proj",
465
+ "layers.9.attn.indexer.compressor.kv_proj",
466
+ "layers.10.attn.compressor.wgate",
467
+ "layers.10.attn.compressor.wkv",
468
+ "layers.10.attn.compressor.fused_wkv_wgate",
469
+ "layers.10.attn.compressor.gate_proj",
470
+ "layers.10.attn.compressor.kv_proj",
471
+ "layers.10.attn.indexer.weights_proj",
472
+ "layers.10.attn.indexer.wq_b",
473
+ "layers.10.attn.indexer.q_b_proj",
474
+ "layers.10.attn.indexer.compressor.wgate",
475
+ "layers.10.attn.indexer.compressor.wkv",
476
+ "layers.10.attn.indexer.compressor.gate_proj",
477
+ "layers.10.attn.indexer.compressor.kv_proj",
478
+ "layers.11.attn.compressor.wgate",
479
+ "layers.11.attn.compressor.wkv",
480
+ "layers.11.attn.compressor.fused_wkv_wgate",
481
+ "layers.11.attn.compressor.gate_proj",
482
+ "layers.11.attn.compressor.kv_proj",
483
+ "layers.11.attn.indexer.weights_proj",
484
+ "layers.11.attn.indexer.wq_b",
485
+ "layers.11.attn.indexer.q_b_proj",
486
+ "layers.11.attn.indexer.compressor.wgate",
487
+ "layers.11.attn.indexer.compressor.wkv",
488
+ "layers.11.attn.indexer.compressor.gate_proj",
489
+ "layers.11.attn.indexer.compressor.kv_proj",
490
+ "layers.12.attn.compressor.wgate",
491
+ "layers.12.attn.compressor.wkv",
492
+ "layers.12.attn.compressor.fused_wkv_wgate",
493
+ "layers.12.attn.compressor.gate_proj",
494
+ "layers.12.attn.compressor.kv_proj",
495
+ "layers.12.attn.indexer.weights_proj",
496
+ "layers.12.attn.indexer.wq_b",
497
+ "layers.12.attn.indexer.q_b_proj",
498
+ "layers.12.attn.indexer.compressor.wgate",
499
+ "layers.12.attn.indexer.compressor.wkv",
500
+ "layers.12.attn.indexer.compressor.gate_proj",
501
+ "layers.12.attn.indexer.compressor.kv_proj",
502
+ "layers.13.attn.compressor.wgate",
503
+ "layers.13.attn.compressor.wkv",
504
+ "layers.13.attn.compressor.fused_wkv_wgate",
505
+ "layers.13.attn.compressor.gate_proj",
506
+ "layers.13.attn.compressor.kv_proj",
507
+ "layers.13.attn.indexer.weights_proj",
508
+ "layers.13.attn.indexer.wq_b",
509
+ "layers.13.attn.indexer.q_b_proj",
510
+ "layers.13.attn.indexer.compressor.wgate",
511
+ "layers.13.attn.indexer.compressor.wkv",
512
+ "layers.13.attn.indexer.compressor.gate_proj",
513
+ "layers.13.attn.indexer.compressor.kv_proj",
514
+ "layers.14.attn.compressor.wgate",
515
+ "layers.14.attn.compressor.wkv",
516
+ "layers.14.attn.compressor.fused_wkv_wgate",
517
+ "layers.14.attn.compressor.gate_proj",
518
+ "layers.14.attn.compressor.kv_proj",
519
+ "layers.14.attn.indexer.weights_proj",
520
+ "layers.14.attn.indexer.wq_b",
521
+ "layers.14.attn.indexer.q_b_proj",
522
+ "layers.14.attn.indexer.compressor.wgate",
523
+ "layers.14.attn.indexer.compressor.wkv",
524
+ "layers.14.attn.indexer.compressor.gate_proj",
525
+ "layers.14.attn.indexer.compressor.kv_proj",
526
+ "layers.15.attn.compressor.wgate",
527
+ "layers.15.attn.compressor.wkv",
528
+ "layers.15.attn.compressor.fused_wkv_wgate",
529
+ "layers.15.attn.compressor.gate_proj",
530
+ "layers.15.attn.compressor.kv_proj",
531
+ "layers.15.attn.indexer.weights_proj",
532
+ "layers.15.attn.indexer.wq_b",
533
+ "layers.15.attn.indexer.q_b_proj",
534
+ "layers.15.attn.indexer.compressor.wgate",
535
+ "layers.15.attn.indexer.compressor.wkv",
536
+ "layers.15.attn.indexer.compressor.gate_proj",
537
+ "layers.15.attn.indexer.compressor.kv_proj",
538
+ "layers.16.attn.compressor.wgate",
539
+ "layers.16.attn.compressor.wkv",
540
+ "layers.16.attn.compressor.fused_wkv_wgate",
541
+ "layers.16.attn.compressor.gate_proj",
542
+ "layers.16.attn.compressor.kv_proj",
543
+ "layers.16.attn.indexer.weights_proj",
544
+ "layers.16.attn.indexer.wq_b",
545
+ "layers.16.attn.indexer.q_b_proj",
546
+ "layers.16.attn.indexer.compressor.wgate",
547
+ "layers.16.attn.indexer.compressor.wkv",
548
+ "layers.16.attn.indexer.compressor.gate_proj",
549
+ "layers.16.attn.indexer.compressor.kv_proj",
550
+ "layers.17.attn.compressor.wgate",
551
+ "layers.17.attn.compressor.wkv",
552
+ "layers.17.attn.compressor.fused_wkv_wgate",
553
+ "layers.17.attn.compressor.gate_proj",
554
+ "layers.17.attn.compressor.kv_proj",
555
+ "layers.17.attn.indexer.weights_proj",
556
+ "layers.17.attn.indexer.wq_b",
557
+ "layers.17.attn.indexer.q_b_proj",
558
+ "layers.17.attn.indexer.compressor.wgate",
559
+ "layers.17.attn.indexer.compressor.wkv",
560
+ "layers.17.attn.indexer.compressor.gate_proj",
561
+ "layers.17.attn.indexer.compressor.kv_proj",
562
+ "layers.18.attn.compressor.wgate",
563
+ "layers.18.attn.compressor.wkv",
564
+ "layers.18.attn.compressor.fused_wkv_wgate",
565
+ "layers.18.attn.compressor.gate_proj",
566
+ "layers.18.attn.compressor.kv_proj",
567
+ "layers.18.attn.indexer.weights_proj",
568
+ "layers.18.attn.indexer.wq_b",
569
+ "layers.18.attn.indexer.q_b_proj",
570
+ "layers.18.attn.indexer.compressor.wgate",
571
+ "layers.18.attn.indexer.compressor.wkv",
572
+ "layers.18.attn.indexer.compressor.gate_proj",
573
+ "layers.18.attn.indexer.compressor.kv_proj",
574
+ "layers.19.attn.compressor.wgate",
575
+ "layers.19.attn.compressor.wkv",
576
+ "layers.19.attn.compressor.fused_wkv_wgate",
577
+ "layers.19.attn.compressor.gate_proj",
578
+ "layers.19.attn.compressor.kv_proj",
579
+ "layers.19.attn.indexer.weights_proj",
580
+ "layers.19.attn.indexer.wq_b",
581
+ "layers.19.attn.indexer.q_b_proj",
582
+ "layers.19.attn.indexer.compressor.wgate",
583
+ "layers.19.attn.indexer.compressor.wkv",
584
+ "layers.19.attn.indexer.compressor.gate_proj",
585
+ "layers.19.attn.indexer.compressor.kv_proj",
586
+ "layers.20.attn.compressor.wgate",
587
+ "layers.20.attn.compressor.wkv",
588
+ "layers.20.attn.compressor.fused_wkv_wgate",
589
+ "layers.20.attn.compressor.gate_proj",
590
+ "layers.20.attn.compressor.kv_proj",
591
+ "layers.20.attn.indexer.weights_proj",
592
+ "layers.20.attn.indexer.wq_b",
593
+ "layers.20.attn.indexer.q_b_proj",
594
+ "layers.20.attn.indexer.compressor.wgate",
595
+ "layers.20.attn.indexer.compressor.wkv",
596
+ "layers.20.attn.indexer.compressor.gate_proj",
597
+ "layers.20.attn.indexer.compressor.kv_proj",
598
+ "layers.21.attn.compressor.wgate",
599
+ "layers.21.attn.compressor.wkv",
600
+ "layers.21.attn.compressor.fused_wkv_wgate",
601
+ "layers.21.attn.compressor.gate_proj",
602
+ "layers.21.attn.compressor.kv_proj",
603
+ "layers.21.attn.indexer.weights_proj",
604
+ "layers.21.attn.indexer.wq_b",
605
+ "layers.21.attn.indexer.q_b_proj",
606
+ "layers.21.attn.indexer.compressor.wgate",
607
+ "layers.21.attn.indexer.compressor.wkv",
608
+ "layers.21.attn.indexer.compressor.gate_proj",
609
+ "layers.21.attn.indexer.compressor.kv_proj",
610
+ "layers.22.attn.compressor.wgate",
611
+ "layers.22.attn.compressor.wkv",
612
+ "layers.22.attn.compressor.fused_wkv_wgate",
613
+ "layers.22.attn.compressor.gate_proj",
614
+ "layers.22.attn.compressor.kv_proj",
615
+ "layers.22.attn.indexer.weights_proj",
616
+ "layers.22.attn.indexer.wq_b",
617
+ "layers.22.attn.indexer.q_b_proj",
618
+ "layers.22.attn.indexer.compressor.wgate",
619
+ "layers.22.attn.indexer.compressor.wkv",
620
+ "layers.22.attn.indexer.compressor.gate_proj",
621
+ "layers.22.attn.indexer.compressor.kv_proj",
622
+ "layers.23.attn.compressor.wgate",
623
+ "layers.23.attn.compressor.wkv",
624
+ "layers.23.attn.compressor.fused_wkv_wgate",
625
+ "layers.23.attn.compressor.gate_proj",
626
+ "layers.23.attn.compressor.kv_proj",
627
+ "layers.23.attn.indexer.weights_proj",
628
+ "layers.23.attn.indexer.wq_b",
629
+ "layers.23.attn.indexer.q_b_proj",
630
+ "layers.23.attn.indexer.compressor.wgate",
631
+ "layers.23.attn.indexer.compressor.wkv",
632
+ "layers.23.attn.indexer.compressor.gate_proj",
633
+ "layers.23.attn.indexer.compressor.kv_proj",
634
+ "layers.24.attn.compressor.wgate",
635
+ "layers.24.attn.compressor.wkv",
636
+ "layers.24.attn.compressor.fused_wkv_wgate",
637
+ "layers.24.attn.compressor.gate_proj",
638
+ "layers.24.attn.compressor.kv_proj",
639
+ "layers.24.attn.indexer.weights_proj",
640
+ "layers.24.attn.indexer.wq_b",
641
+ "layers.24.attn.indexer.q_b_proj",
642
+ "layers.24.attn.indexer.compressor.wgate",
643
+ "layers.24.attn.indexer.compressor.wkv",
644
+ "layers.24.attn.indexer.compressor.gate_proj",
645
+ "layers.24.attn.indexer.compressor.kv_proj",
646
+ "layers.25.attn.compressor.wgate",
647
+ "layers.25.attn.compressor.wkv",
648
+ "layers.25.attn.compressor.fused_wkv_wgate",
649
+ "layers.25.attn.compressor.gate_proj",
650
+ "layers.25.attn.compressor.kv_proj",
651
+ "layers.25.attn.indexer.weights_proj",
652
+ "layers.25.attn.indexer.wq_b",
653
+ "layers.25.attn.indexer.q_b_proj",
654
+ "layers.25.attn.indexer.compressor.wgate",
655
+ "layers.25.attn.indexer.compressor.wkv",
656
+ "layers.25.attn.indexer.compressor.gate_proj",
657
+ "layers.25.attn.indexer.compressor.kv_proj",
658
+ "layers.26.attn.compressor.wgate",
659
+ "layers.26.attn.compressor.wkv",
660
+ "layers.26.attn.compressor.fused_wkv_wgate",
661
+ "layers.26.attn.compressor.gate_proj",
662
+ "layers.26.attn.compressor.kv_proj",
663
+ "layers.26.attn.indexer.weights_proj",
664
+ "layers.26.attn.indexer.wq_b",
665
+ "layers.26.attn.indexer.q_b_proj",
666
+ "layers.26.attn.indexer.compressor.wgate",
667
+ "layers.26.attn.indexer.compressor.wkv",
668
+ "layers.26.attn.indexer.compressor.gate_proj",
669
+ "layers.26.attn.indexer.compressor.kv_proj",
670
+ "layers.27.attn.compressor.wgate",
671
+ "layers.27.attn.compressor.wkv",
672
+ "layers.27.attn.compressor.fused_wkv_wgate",
673
+ "layers.27.attn.compressor.gate_proj",
674
+ "layers.27.attn.compressor.kv_proj",
675
+ "layers.27.attn.indexer.weights_proj",
676
+ "layers.27.attn.indexer.wq_b",
677
+ "layers.27.attn.indexer.q_b_proj",
678
+ "layers.27.attn.indexer.compressor.wgate",
679
+ "layers.27.attn.indexer.compressor.wkv",
680
+ "layers.27.attn.indexer.compressor.gate_proj",
681
+ "layers.27.attn.indexer.compressor.kv_proj",
682
+ "layers.28.attn.compressor.wgate",
683
+ "layers.28.attn.compressor.wkv",
684
+ "layers.28.attn.compressor.fused_wkv_wgate",
685
+ "layers.28.attn.compressor.gate_proj",
686
+ "layers.28.attn.compressor.kv_proj",
687
+ "layers.28.attn.indexer.weights_proj",
688
+ "layers.28.attn.indexer.wq_b",
689
+ "layers.28.attn.indexer.q_b_proj",
690
+ "layers.28.attn.indexer.compressor.wgate",
691
+ "layers.28.attn.indexer.compressor.wkv",
692
+ "layers.28.attn.indexer.compressor.gate_proj",
693
+ "layers.28.attn.indexer.compressor.kv_proj",
694
+ "layers.29.attn.compressor.wgate",
695
+ "layers.29.attn.compressor.wkv",
696
+ "layers.29.attn.compressor.fused_wkv_wgate",
697
+ "layers.29.attn.compressor.gate_proj",
698
+ "layers.29.attn.compressor.kv_proj",
699
+ "layers.29.attn.indexer.weights_proj",
700
+ "layers.29.attn.indexer.wq_b",
701
+ "layers.29.attn.indexer.q_b_proj",
702
+ "layers.29.attn.indexer.compressor.wgate",
703
+ "layers.29.attn.indexer.compressor.wkv",
704
+ "layers.29.attn.indexer.compressor.gate_proj",
705
+ "layers.29.attn.indexer.compressor.kv_proj",
706
+ "layers.30.attn.compressor.wgate",
707
+ "layers.30.attn.compressor.wkv",
708
+ "layers.30.attn.compressor.fused_wkv_wgate",
709
+ "layers.30.attn.compressor.gate_proj",
710
+ "layers.30.attn.compressor.kv_proj",
711
+ "layers.30.attn.indexer.weights_proj",
712
+ "layers.30.attn.indexer.wq_b",
713
+ "layers.30.attn.indexer.q_b_proj",
714
+ "layers.30.attn.indexer.compressor.wgate",
715
+ "layers.30.attn.indexer.compressor.wkv",
716
+ "layers.30.attn.indexer.compressor.gate_proj",
717
+ "layers.30.attn.indexer.compressor.kv_proj",
718
+ "layers.31.attn.compressor.wgate",
719
+ "layers.31.attn.compressor.wkv",
720
+ "layers.31.attn.compressor.fused_wkv_wgate",
721
+ "layers.31.attn.compressor.gate_proj",
722
+ "layers.31.attn.compressor.kv_proj",
723
+ "layers.31.attn.indexer.weights_proj",
724
+ "layers.31.attn.indexer.wq_b",
725
+ "layers.31.attn.indexer.q_b_proj",
726
+ "layers.31.attn.indexer.compressor.wgate",
727
+ "layers.31.attn.indexer.compressor.wkv",
728
+ "layers.31.attn.indexer.compressor.gate_proj",
729
+ "layers.31.attn.indexer.compressor.kv_proj",
730
+ "layers.32.attn.compressor.wgate",
731
+ "layers.32.attn.compressor.wkv",
732
+ "layers.32.attn.compressor.fused_wkv_wgate",
733
+ "layers.32.attn.compressor.gate_proj",
734
+ "layers.32.attn.compressor.kv_proj",
735
+ "layers.32.attn.indexer.weights_proj",
736
+ "layers.32.attn.indexer.wq_b",
737
+ "layers.32.attn.indexer.q_b_proj",
738
+ "layers.32.attn.indexer.compressor.wgate",
739
+ "layers.32.attn.indexer.compressor.wkv",
740
+ "layers.32.attn.indexer.compressor.gate_proj",
741
+ "layers.32.attn.indexer.compressor.kv_proj",
742
+ "layers.33.attn.compressor.wgate",
743
+ "layers.33.attn.compressor.wkv",
744
+ "layers.33.attn.compressor.fused_wkv_wgate",
745
+ "layers.33.attn.compressor.gate_proj",
746
+ "layers.33.attn.compressor.kv_proj",
747
+ "layers.33.attn.indexer.weights_proj",
748
+ "layers.33.attn.indexer.wq_b",
749
+ "layers.33.attn.indexer.q_b_proj",
750
+ "layers.33.attn.indexer.compressor.wgate",
751
+ "layers.33.attn.indexer.compressor.wkv",
752
+ "layers.33.attn.indexer.compressor.gate_proj",
753
+ "layers.33.attn.indexer.compressor.kv_proj",
754
+ "layers.34.attn.compressor.wgate",
755
+ "layers.34.attn.compressor.wkv",
756
+ "layers.34.attn.compressor.fused_wkv_wgate",
757
+ "layers.34.attn.compressor.gate_proj",
758
+ "layers.34.attn.compressor.kv_proj",
759
+ "layers.34.attn.indexer.weights_proj",
760
+ "layers.34.attn.indexer.wq_b",
761
+ "layers.34.attn.indexer.q_b_proj",
762
+ "layers.34.attn.indexer.compressor.wgate",
763
+ "layers.34.attn.indexer.compressor.wkv",
764
+ "layers.34.attn.indexer.compressor.gate_proj",
765
+ "layers.34.attn.indexer.compressor.kv_proj",
766
+ "layers.35.attn.compressor.wgate",
767
+ "layers.35.attn.compressor.wkv",
768
+ "layers.35.attn.compressor.fused_wkv_wgate",
769
+ "layers.35.attn.compressor.gate_proj",
770
+ "layers.35.attn.compressor.kv_proj",
771
+ "layers.35.attn.indexer.weights_proj",
772
+ "layers.35.attn.indexer.wq_b",
773
+ "layers.35.attn.indexer.q_b_proj",
774
+ "layers.35.attn.indexer.compressor.wgate",
775
+ "layers.35.attn.indexer.compressor.wkv",
776
+ "layers.35.attn.indexer.compressor.gate_proj",
777
+ "layers.35.attn.indexer.compressor.kv_proj",
778
+ "layers.36.attn.compressor.wgate",
779
+ "layers.36.attn.compressor.wkv",
780
+ "layers.36.attn.compressor.fused_wkv_wgate",
781
+ "layers.36.attn.compressor.gate_proj",
782
+ "layers.36.attn.compressor.kv_proj",
783
+ "layers.36.attn.indexer.weights_proj",
784
+ "layers.36.attn.indexer.wq_b",
785
+ "layers.36.attn.indexer.q_b_proj",
786
+ "layers.36.attn.indexer.compressor.wgate",
787
+ "layers.36.attn.indexer.compressor.wkv",
788
+ "layers.36.attn.indexer.compressor.gate_proj",
789
+ "layers.36.attn.indexer.compressor.kv_proj",
790
+ "layers.37.attn.compressor.wgate",
791
+ "layers.37.attn.compressor.wkv",
792
+ "layers.37.attn.compressor.fused_wkv_wgate",
793
+ "layers.37.attn.compressor.gate_proj",
794
+ "layers.37.attn.compressor.kv_proj",
795
+ "layers.37.attn.indexer.weights_proj",
796
+ "layers.37.attn.indexer.wq_b",
797
+ "layers.37.attn.indexer.q_b_proj",
798
+ "layers.37.attn.indexer.compressor.wgate",
799
+ "layers.37.attn.indexer.compressor.wkv",
800
+ "layers.37.attn.indexer.compressor.gate_proj",
801
+ "layers.37.attn.indexer.compressor.kv_proj",
802
+ "layers.38.attn.compressor.wgate",
803
+ "layers.38.attn.compressor.wkv",
804
+ "layers.38.attn.compressor.fused_wkv_wgate",
805
+ "layers.38.attn.compressor.gate_proj",
806
+ "layers.38.attn.compressor.kv_proj",
807
+ "layers.38.attn.indexer.weights_proj",
808
+ "layers.38.attn.indexer.wq_b",
809
+ "layers.38.attn.indexer.q_b_proj",
810
+ "layers.38.attn.indexer.compressor.wgate",
811
+ "layers.38.attn.indexer.compressor.wkv",
812
+ "layers.38.attn.indexer.compressor.gate_proj",
813
+ "layers.38.attn.indexer.compressor.kv_proj",
814
+ "layers.39.attn.compressor.wgate",
815
+ "layers.39.attn.compressor.wkv",
816
+ "layers.39.attn.compressor.fused_wkv_wgate",
817
+ "layers.39.attn.compressor.gate_proj",
818
+ "layers.39.attn.compressor.kv_proj",
819
+ "layers.39.attn.indexer.weights_proj",
820
+ "layers.39.attn.indexer.wq_b",
821
+ "layers.39.attn.indexer.q_b_proj",
822
+ "layers.39.attn.indexer.compressor.wgate",
823
+ "layers.39.attn.indexer.compressor.wkv",
824
+ "layers.39.attn.indexer.compressor.gate_proj",
825
+ "layers.39.attn.indexer.compressor.kv_proj",
826
+ "layers.40.attn.compressor.wgate",
827
+ "layers.40.attn.compressor.wkv",
828
+ "layers.40.attn.compressor.fused_wkv_wgate",
829
+ "layers.40.attn.compressor.gate_proj",
830
+ "layers.40.attn.compressor.kv_proj",
831
+ "layers.40.attn.indexer.weights_proj",
832
+ "layers.40.attn.indexer.wq_b",
833
+ "layers.40.attn.indexer.q_b_proj",
834
+ "layers.40.attn.indexer.compressor.wgate",
835
+ "layers.40.attn.indexer.compressor.wkv",
836
+ "layers.40.attn.indexer.compressor.gate_proj",
837
+ "layers.40.attn.indexer.compressor.kv_proj",
838
+ "layers.41.attn.compressor.wgate",
839
+ "layers.41.attn.compressor.wkv",
840
+ "layers.41.attn.compressor.fused_wkv_wgate",
841
+ "layers.41.attn.compressor.gate_proj",
842
+ "layers.41.attn.compressor.kv_proj",
843
+ "layers.41.attn.indexer.weights_proj",
844
+ "layers.41.attn.indexer.wq_b",
845
+ "layers.41.attn.indexer.q_b_proj",
846
+ "layers.41.attn.indexer.compressor.wgate",
847
+ "layers.41.attn.indexer.compressor.wkv",
848
+ "layers.41.attn.indexer.compressor.gate_proj",
849
+ "layers.41.attn.indexer.compressor.kv_proj",
850
+ "layers.42.attn.compressor.wgate",
851
+ "layers.42.attn.compressor.wkv",
852
+ "layers.42.attn.compressor.fused_wkv_wgate",
853
+ "layers.42.attn.compressor.gate_proj",
854
+ "layers.42.attn.compressor.kv_proj",
855
+ "layers.42.attn.indexer.weights_proj",
856
+ "layers.42.attn.indexer.wq_b",
857
+ "layers.42.attn.indexer.q_b_proj",
858
+ "layers.42.attn.indexer.compressor.wgate",
859
+ "layers.42.attn.indexer.compressor.wkv",
860
+ "layers.42.attn.indexer.compressor.gate_proj",
861
+ "layers.42.attn.indexer.compressor.kv_proj",
862
+ "layers.43.attn.compressor.wgate",
863
+ "layers.43.attn.compressor.wkv",
864
+ "layers.43.attn.compressor.fused_wkv_wgate",
865
+ "layers.43.attn.compressor.gate_proj",
866
+ "layers.43.attn.compressor.kv_proj",
867
+ "layers.43.attn.indexer.weights_proj",
868
+ "layers.43.attn.indexer.wq_b",
869
+ "layers.43.attn.indexer.q_b_proj",
870
+ "layers.43.attn.indexer.compressor.wgate",
871
+ "layers.43.attn.indexer.compressor.wkv",
872
+ "layers.43.attn.indexer.compressor.gate_proj",
873
+ "layers.43.attn.indexer.compressor.kv_proj",
874
+ "layers.44.attn.compressor.wgate",
875
+ "layers.44.attn.compressor.wkv",
876
+ "layers.44.attn.compressor.fused_wkv_wgate",
877
+ "layers.44.attn.compressor.gate_proj",
878
+ "layers.44.attn.compressor.kv_proj",
879
+ "layers.44.attn.indexer.weights_proj",
880
+ "layers.44.attn.indexer.wq_b",
881
+ "layers.44.attn.indexer.q_b_proj",
882
+ "layers.44.attn.indexer.compressor.wgate",
883
+ "layers.44.attn.indexer.compressor.wkv",
884
+ "layers.44.attn.indexer.compressor.gate_proj",
885
+ "layers.44.attn.indexer.compressor.kv_proj",
886
+ "layers.45.attn.compressor.wgate",
887
+ "layers.45.attn.compressor.wkv",
888
+ "layers.45.attn.compressor.fused_wkv_wgate",
889
+ "layers.45.attn.compressor.gate_proj",
890
+ "layers.45.attn.compressor.kv_proj",
891
+ "layers.45.attn.indexer.weights_proj",
892
+ "layers.45.attn.indexer.wq_b",
893
+ "layers.45.attn.indexer.q_b_proj",
894
+ "layers.45.attn.indexer.compressor.wgate",
895
+ "layers.45.attn.indexer.compressor.wkv",
896
+ "layers.45.attn.indexer.compressor.gate_proj",
897
+ "layers.45.attn.indexer.compressor.kv_proj",
898
+ "layers.46.attn.compressor.wgate",
899
+ "layers.46.attn.compressor.wkv",
900
+ "layers.46.attn.compressor.fused_wkv_wgate",
901
+ "layers.46.attn.compressor.gate_proj",
902
+ "layers.46.attn.compressor.kv_proj",
903
+ "layers.46.attn.indexer.weights_proj",
904
+ "layers.46.attn.indexer.wq_b",
905
+ "layers.46.attn.indexer.q_b_proj",
906
+ "layers.46.attn.indexer.compressor.wgate",
907
+ "layers.46.attn.indexer.compressor.wkv",
908
+ "layers.46.attn.indexer.compressor.gate_proj",
909
+ "layers.46.attn.indexer.compressor.kv_proj",
910
+ "layers.47.attn.compressor.wgate",
911
+ "layers.47.attn.compressor.wkv",
912
+ "layers.47.attn.compressor.fused_wkv_wgate",
913
+ "layers.47.attn.compressor.gate_proj",
914
+ "layers.47.attn.compressor.kv_proj",
915
+ "layers.47.attn.indexer.weights_proj",
916
+ "layers.47.attn.indexer.wq_b",
917
+ "layers.47.attn.indexer.q_b_proj",
918
+ "layers.47.attn.indexer.compressor.wgate",
919
+ "layers.47.attn.indexer.compressor.wkv",
920
+ "layers.47.attn.indexer.compressor.gate_proj",
921
+ "layers.47.attn.indexer.compressor.kv_proj",
922
+ "layers.48.attn.compressor.wgate",
923
+ "layers.48.attn.compressor.wkv",
924
+ "layers.48.attn.compressor.fused_wkv_wgate",
925
+ "layers.48.attn.compressor.gate_proj",
926
+ "layers.48.attn.compressor.kv_proj",
927
+ "layers.48.attn.indexer.weights_proj",
928
+ "layers.48.attn.indexer.wq_b",
929
+ "layers.48.attn.indexer.q_b_proj",
930
+ "layers.48.attn.indexer.compressor.wgate",
931
+ "layers.48.attn.indexer.compressor.wkv",
932
+ "layers.48.attn.indexer.compressor.gate_proj",
933
+ "layers.48.attn.indexer.compressor.kv_proj",
934
+ "layers.49.attn.compressor.wgate",
935
+ "layers.49.attn.compressor.wkv",
936
+ "layers.49.attn.compressor.fused_wkv_wgate",
937
+ "layers.49.attn.compressor.gate_proj",
938
+ "layers.49.attn.compressor.kv_proj",
939
+ "layers.49.attn.indexer.weights_proj",
940
+ "layers.49.attn.indexer.wq_b",
941
+ "layers.49.attn.indexer.q_b_proj",
942
+ "layers.49.attn.indexer.compressor.wgate",
943
+ "layers.49.attn.indexer.compressor.wkv",
944
+ "layers.49.attn.indexer.compressor.gate_proj",
945
+ "layers.49.attn.indexer.compressor.kv_proj",
946
+ "layers.50.attn.compressor.wgate",
947
+ "layers.50.attn.compressor.wkv",
948
+ "layers.50.attn.compressor.fused_wkv_wgate",
949
+ "layers.50.attn.compressor.gate_proj",
950
+ "layers.50.attn.compressor.kv_proj",
951
+ "layers.50.attn.indexer.weights_proj",
952
+ "layers.50.attn.indexer.wq_b",
953
+ "layers.50.attn.indexer.q_b_proj",
954
+ "layers.50.attn.indexer.compressor.wgate",
955
+ "layers.50.attn.indexer.compressor.wkv",
956
+ "layers.50.attn.indexer.compressor.gate_proj",
957
+ "layers.50.attn.indexer.compressor.kv_proj",
958
+ "layers.51.attn.compressor.wgate",
959
+ "layers.51.attn.compressor.wkv",
960
+ "layers.51.attn.compressor.fused_wkv_wgate",
961
+ "layers.51.attn.compressor.gate_proj",
962
+ "layers.51.attn.compressor.kv_proj",
963
+ "layers.51.attn.indexer.weights_proj",
964
+ "layers.51.attn.indexer.wq_b",
965
+ "layers.51.attn.indexer.q_b_proj",
966
+ "layers.51.attn.indexer.compressor.wgate",
967
+ "layers.51.attn.indexer.compressor.wkv",
968
+ "layers.51.attn.indexer.compressor.gate_proj",
969
+ "layers.51.attn.indexer.compressor.kv_proj",
970
+ "layers.52.attn.compressor.wgate",
971
+ "layers.52.attn.compressor.wkv",
972
+ "layers.52.attn.compressor.fused_wkv_wgate",
973
+ "layers.52.attn.compressor.gate_proj",
974
+ "layers.52.attn.compressor.kv_proj",
975
+ "layers.52.attn.indexer.weights_proj",
976
+ "layers.52.attn.indexer.wq_b",
977
+ "layers.52.attn.indexer.q_b_proj",
978
+ "layers.52.attn.indexer.compressor.wgate",
979
+ "layers.52.attn.indexer.compressor.wkv",
980
+ "layers.52.attn.indexer.compressor.gate_proj",
981
+ "layers.52.attn.indexer.compressor.kv_proj",
982
+ "layers.53.attn.compressor.wgate",
983
+ "layers.53.attn.compressor.wkv",
984
+ "layers.53.attn.compressor.fused_wkv_wgate",
985
+ "layers.53.attn.compressor.gate_proj",
986
+ "layers.53.attn.compressor.kv_proj",
987
+ "layers.53.attn.indexer.weights_proj",
988
+ "layers.53.attn.indexer.wq_b",
989
+ "layers.53.attn.indexer.q_b_proj",
990
+ "layers.53.attn.indexer.compressor.wgate",
991
+ "layers.53.attn.indexer.compressor.wkv",
992
+ "layers.53.attn.indexer.compressor.gate_proj",
993
+ "layers.53.attn.indexer.compressor.kv_proj",
994
+ "layers.54.attn.compressor.wgate",
995
+ "layers.54.attn.compressor.wkv",
996
+ "layers.54.attn.compressor.fused_wkv_wgate",
997
+ "layers.54.attn.compressor.gate_proj",
998
+ "layers.54.attn.compressor.kv_proj",
999
+ "layers.54.attn.indexer.weights_proj",
1000
+ "layers.54.attn.indexer.wq_b",
1001
+ "layers.54.attn.indexer.q_b_proj",
1002
+ "layers.54.attn.indexer.compressor.wgate",
1003
+ "layers.54.attn.indexer.compressor.wkv",
1004
+ "layers.54.attn.indexer.compressor.gate_proj",
1005
+ "layers.54.attn.indexer.compressor.kv_proj",
1006
+ "layers.55.attn.compressor.wgate",
1007
+ "layers.55.attn.compressor.wkv",
1008
+ "layers.55.attn.compressor.fused_wkv_wgate",
1009
+ "layers.55.attn.compressor.gate_proj",
1010
+ "layers.55.attn.compressor.kv_proj",
1011
+ "layers.55.attn.indexer.weights_proj",
1012
+ "layers.55.attn.indexer.wq_b",
1013
+ "layers.55.attn.indexer.q_b_proj",
1014
+ "layers.55.attn.indexer.compressor.wgate",
1015
+ "layers.55.attn.indexer.compressor.wkv",
1016
+ "layers.55.attn.indexer.compressor.gate_proj",
1017
+ "layers.55.attn.indexer.compressor.kv_proj",
1018
+ "layers.56.attn.compressor.wgate",
1019
+ "layers.56.attn.compressor.wkv",
1020
+ "layers.56.attn.compressor.fused_wkv_wgate",
1021
+ "layers.56.attn.compressor.gate_proj",
1022
+ "layers.56.attn.compressor.kv_proj",
1023
+ "layers.56.attn.indexer.weights_proj",
1024
+ "layers.56.attn.indexer.wq_b",
1025
+ "layers.56.attn.indexer.q_b_proj",
1026
+ "layers.56.attn.indexer.compressor.wgate",
1027
+ "layers.56.attn.indexer.compressor.wkv",
1028
+ "layers.56.attn.indexer.compressor.gate_proj",
1029
+ "layers.56.attn.indexer.compressor.kv_proj",
1030
+ "layers.57.attn.compressor.wgate",
1031
+ "layers.57.attn.compressor.wkv",
1032
+ "layers.57.attn.compressor.fused_wkv_wgate",
1033
+ "layers.57.attn.compressor.gate_proj",
1034
+ "layers.57.attn.compressor.kv_proj",
1035
+ "layers.57.attn.indexer.weights_proj",
1036
+ "layers.57.attn.indexer.wq_b",
1037
+ "layers.57.attn.indexer.q_b_proj",
1038
+ "layers.57.attn.indexer.compressor.wgate",
1039
+ "layers.57.attn.indexer.compressor.wkv",
1040
+ "layers.57.attn.indexer.compressor.gate_proj",
1041
+ "layers.57.attn.indexer.compressor.kv_proj",
1042
+ "layers.58.attn.compressor.wgate",
1043
+ "layers.58.attn.compressor.wkv",
1044
+ "layers.58.attn.compressor.fused_wkv_wgate",
1045
+ "layers.58.attn.compressor.gate_proj",
1046
+ "layers.58.attn.compressor.kv_proj",
1047
+ "layers.58.attn.indexer.weights_proj",
1048
+ "layers.58.attn.indexer.wq_b",
1049
+ "layers.58.attn.indexer.q_b_proj",
1050
+ "layers.58.attn.indexer.compressor.wgate",
1051
+ "layers.58.attn.indexer.compressor.wkv",
1052
+ "layers.58.attn.indexer.compressor.gate_proj",
1053
+ "layers.58.attn.indexer.compressor.kv_proj",
1054
+ "layers.59.attn.compressor.wgate",
1055
+ "layers.59.attn.compressor.wkv",
1056
+ "layers.59.attn.compressor.fused_wkv_wgate",
1057
+ "layers.59.attn.compressor.gate_proj",
1058
+ "layers.59.attn.compressor.kv_proj",
1059
+ "layers.59.attn.indexer.weights_proj",
1060
+ "layers.59.attn.indexer.wq_b",
1061
+ "layers.59.attn.indexer.q_b_proj",
1062
+ "layers.59.attn.indexer.compressor.wgate",
1063
+ "layers.59.attn.indexer.compressor.wkv",
1064
+ "layers.59.attn.indexer.compressor.gate_proj",
1065
+ "layers.59.attn.indexer.compressor.kv_proj",
1066
+ "layers.60.attn.compressor.wgate",
1067
+ "layers.60.attn.compressor.wkv",
1068
+ "layers.60.attn.compressor.fused_wkv_wgate",
1069
+ "layers.60.attn.compressor.gate_proj",
1070
+ "layers.60.attn.compressor.kv_proj",
1071
+ "layers.60.attn.indexer.weights_proj",
1072
+ "layers.60.attn.indexer.wq_b",
1073
+ "layers.60.attn.indexer.q_b_proj",
1074
+ "layers.60.attn.indexer.compressor.wgate",
1075
+ "layers.60.attn.indexer.compressor.wkv",
1076
+ "layers.60.attn.indexer.compressor.gate_proj",
1077
+ "layers.60.attn.indexer.compressor.kv_proj"
1078
+ ],
1079
+ "kv_cache_scheme": null,
1080
+ "quant_method": "compressed-tensors",
1081
+ "quantization_status": "compressed",
1082
+ "sparsity_config": {},
1083
+ "transform_config": {},
1084
+ "version": "0.15.1.a20260515",
1085
+ "scale_fmt": "ue8m0"
1086
+ },
1087
+ "rms_norm_eps": 1e-06,
1088
+ "rope_parameters": {
1089
+ "compress": {
1090
+ "attention_factor": 1.0,
1091
+ "beta_fast": 32,
1092
+ "beta_slow": 1,
1093
+ "factor": 16,
1094
+ "original_max_position_embeddings": 65536,
1095
+ "partial_rotary_factor": 0.125,
1096
+ "rope_theta": 160000,
1097
+ "rope_type": "yarn",
1098
+ "type": "yarn"
1099
+ },
1100
+ "main": {
1101
+ "partial_rotary_factor": 0.125,
1102
+ "rope_theta": 10000,
1103
+ "rope_type": "default"
1104
+ },
1105
+ "partial_rotary_factor": 0.125,
1106
+ "rope_theta": 10000,
1107
+ "rope_type": "default"
1108
+ },
1109
+ "rope_theta": 10000,
1110
+ "routed_scaling_factor": 1.5,
1111
+ "router_aux_loss_coef": 0.001,
1112
+ "router_jitter_noise": 0.0,
1113
+ "scoring_func": "sqrtsoftplus",
1114
+ "sliding_window": 128,
1115
+ "swiglu_limit": 10.0,
1116
+ "tie_word_embeddings": false,
1117
+ "topk_method": "noaux_tc",
1118
+ "transformers_version": "5.8.1",
1119
+ "use_cache": true,
1120
+ "vocab_size": 129280,
1121
+ "compress_ratios": [
1122
+ 0,
1123
+ 0,
1124
+ 4,
1125
+ 128,
1126
+ 4,
1127
+ 128,
1128
+ 4,
1129
+ 128,
1130
+ 4,
1131
+ 128,
1132
+ 4,
1133
+ 128,
1134
+ 4,
1135
+ 128,
1136
+ 4,
1137
+ 128,
1138
+ 4,
1139
+ 128,
1140
+ 4,
1141
+ 128,
1142
+ 4,
1143
+ 128,
1144
+ 4,
1145
+ 128,
1146
+ 4,
1147
+ 128,
1148
+ 4,
1149
+ 128,
1150
+ 4,
1151
+ 128,
1152
+ 4,
1153
+ 128,
1154
+ 4,
1155
+ 128,
1156
+ 4,
1157
+ 128,
1158
+ 4,
1159
+ 128,
1160
+ 4,
1161
+ 128,
1162
+ 4,
1163
+ 128,
1164
+ 4,
1165
+ 0
1166
+ ],
1167
+ "num_hash_layers": 3,
1168
+ "rope_scaling": {
1169
+ "beta_fast": 32,
1170
+ "beta_slow": 1,
1171
+ "factor": 16,
1172
+ "original_max_position_embeddings": 65536,
1173
+ "type": "yarn"
1174
+ },
1175
+ "torch_dtype": "bfloat16",
1176
+ "_w4a8_conversion": "experts input_activations: null -> fp8 dynamic token (Claude/Evrard 2026-06-26)"
1177
+ }
generation_config.json ADDED
@@ -0,0 +1,9 @@
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "_from_model_config": true,
3
+ "bos_token_id": 0,
4
+ "do_sample": true,
5
+ "eos_token_id": 1,
6
+ "temperature": 1.0,
7
+ "top_p": 1.0,
8
+ "transformers_version": "5.8.1"
9
+ }
model.safetensors.index.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer.json ADDED
The diff for this file is too large to render. See raw diff
 
tokenizer_config.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "backend": "tokenizers",
3
+ "bos_token": "<｜begin▁of▁sentence｜>",
4
+ "clean_up_tokenization_spaces": false,
5
+ "eos_token": "<｜end▁of▁sentence｜>",
6
+ "is_local": true,
7
+ "legacy": true,
8
+ "local_files_only": false,
9
+ "model_max_length": 1048576,
10
+ "pad_token": "<｜end▁of▁sentence｜>",
11
+ "sp_model_kwargs": {},
12
+ "tokenizer_class": "TokenizersBackend",
13
+ "unk_token": null
14
+ }