hanxiao commited on
Commit
abec4c4
·
verified ·
1 Parent(s): 425faa5

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +52 -23
README.md CHANGED
@@ -120,44 +120,73 @@ Here's the speed and quality evaluation on two nano benchmarks. The higher the b
120
  ![](https://raw.githubusercontent.com/jina-ai/jina-embeddings-v4-gguf/refs/heads/main/NanoHotpotQA.svg)
121
  ![](https://raw.githubusercontent.com/jina-ai/jina-embeddings-v4-gguf/refs/heads/main/NanoFiQA2018.svg)
122
 
123
- #### Table 1: Tokens per Second
124
-
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
125
  ```
 
 
 
 
 
 
 
 
126
  llama_context: n_seq_max = 1
127
  llama_context: n_ctx = 4096
128
  llama_context: n_ctx_per_seq = 4096
129
  llama_context: n_batch = 4096
130
  llama_context: n_ubatch = 4096
131
  llama_context: causal_attn = 1
132
- llama_context: flash_attn = 0
133
  llama_context: kv_unified = true
134
  llama_context: freq_base = 1000000.0
135
  llama_context: freq_scale = 1
136
  llama_context: n_ctx_per_seq (4096) < n_ctx_train (128000) -- the full capacity of the model will not be utilized
 
 
 
 
 
 
 
 
 
 
 
 
 
 
137
 
138
  system_info: n_threads = 4 (n_threads_batch = 4) / 8 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
 
 
139
  ```
140
 
141
- | Quantization Type | File Size | BPW | NanoHotpotQA | NanoFiQA2018 | Δ to F16 (HotpotQA) | Δ to F16 (FiQA2018) |
142
- |------------------|-----------|-----|--------------|--------------|--------------------|--------------------|
143
- | IQ1_S | 748.77 MiB | 2.04 | 1608 | 1618 | +53% | +49% |
144
- | IQ1_M | 804.97 MiB | 2.19 | 1553 | 1563 | +48% | +44% |
145
- | IQ2_XXS | 898.64 MiB | 2.44 | 1600 | 1612 | +52% | +49% |
146
- | IQ2_M | 1.06 GiB | 2.94 | 1529 | 1534 | +46% | +42% |
147
- | Q2_K | 1.18 GiB | 3.29 | 1459 | 1471 | +39% | +36% |
148
- | IQ3_XXS | 1.19 GiB | 3.31 | 1552 | 1487 | +48% | +37% |
149
- | IQ3_XS | 1.29 GiB | 3.59 | 1529 | 1526 | +46% | +41% |
150
- | IQ3_S | 1.35 GiB | 3.76 | 1520 | 1516 | +45% | +40% |
151
- | IQ3_M | 1.38 GiB | 3.84 | 1507 | 1511 | +44% | +40% |
152
- | Q3_K_M | 1.48 GiB | 4.11 | 1475 | 1487 | +40% | +37% |
153
- | IQ4_NL | 1.69 GiB | 4.72 | 1464 | 1469 | +39% | +36% |
154
- | IQ4_XS | 1.61 GiB | 4.49 | 1478 | 1487 | +41% | +37% |
155
- | Q4_K_M | 1.79 GiB | 4.99 | 1454 | 1458 | +38% | +35% |
156
- | Q5_K_S | 2.02 GiB | 5.61 | 1419 | 1429 | +35% | +32% |
157
- | Q5_K_M | 2.07 GiB | 5.75 | 1404 | 1433 | +34% | +32% |
158
- | Q6_K | 2.36 GiB | 6.56 | 1356 | 1382 | +29% | +28% |
159
- | Q8_0 | 3.05 GiB | 8.50 | 1304 | 1334 | +24% | +23% |
160
- | F16 | 5.75 GiB | 16.00 | 1050 | 1083 | +0% | +0% |
161
 
162
  #### Table 2: NDCG@5
163
  ## NDCG@5 Performance Comparison
 
120
  ![](https://raw.githubusercontent.com/jina-ai/jina-embeddings-v4-gguf/refs/heads/main/NanoHotpotQA.svg)
121
  ![](https://raw.githubusercontent.com/jina-ai/jina-embeddings-v4-gguf/refs/heads/main/NanoFiQA2018.svg)
122
 
123
+ #### Table 1: Tokens per Second on NanoHotpotQA Documents
124
+
125
+
126
+ | Quantization Type | File Size | BPW | Peak VRAM | Token/s | Δ to F16 |
127
+ |------------------|-----------|-----|-----------|--------------|----------|
128
+ | IQ1_S | 748.77 MiB | 2.04 | 3651MB | 3625 | +7% |
129
+ | IQ1_M | 804.97 MiB | 2.19 | 3799MB | 3349 | -1% |
130
+ | IQ2_XXS | 898.64 MiB | 2.44 | 3799MB | 3701 | +9% |
131
+ | IQ2_M | 1.06 GiB | 2.94 | 3983MB | 3407 | +0% |
132
+ | Q2_K | 1.18 GiB | 3.29 | 4113MB | 3173 | -7% |
133
+ | IQ3_XXS | 1.19 GiB | 3.31 | 4119MB | 3668 | +8% |
134
+ | IQ3_XS | 1.29 GiB | 3.59 | 4221MB | 3604 | +6% |
135
+ | IQ3_S | 1.35 GiB | 3.76 | 4283MB | 3599 | +6% |
136
+ | IQ3_M | 1.38 GiB | 3.84 | 4315MB | 3603 | +6% |
137
+ | Q3_K_M | 1.48 GiB | 4.11 | 4411MB | 3450 | +2% |
138
+ | IQ4_NL | 1.69 GiB | 4.72 | 4635MB | 3571 | +5% |
139
+ | IQ4_XS | 1.61 GiB | 4.49 | 4553MB | 3585 | +5% |
140
+ | Q4_K_M | 1.79 GiB | 4.99 | 4735MB | 3558 | +5% |
141
+ | Q5_K_S | 2.02 GiB | 5.61 | 4963MB | 3567 | +5% |
142
+ | Q5_K_M | 2.07 GiB | 5.75 | 5017MB | 3528 | +4% |
143
+ | Q6_K | 2.36 GiB | 6.56 | 5315MB | 3334 | -2% |
144
+ | Q8_0 | 3.05 GiB | 8.50 | 6027MB | 3767 | +11% |
145
+ | F16 | 5.75 GiB | 16.00 | 9939MB | 3399 | +0% |
146
+
147
+
148
+ System info:
149
  ```
150
+ load_tensors: loading model tensors, this can take a while... (mmap = true)
151
+ load_tensors: offloading 36 repeating layers to GPU
152
+ load_tensors: offloading output layer to GPU
153
+ load_tensors: offloaded 37/37 layers to GPU
154
+ load_tensors: CUDA0 model buffer size = 3127.61 MiB
155
+ load_tensors: CPU_Mapped model buffer size = 315.30 MiB
156
+ ...................................................................................
157
+ llama_context: constructing llama_context
158
  llama_context: n_seq_max = 1
159
  llama_context: n_ctx = 4096
160
  llama_context: n_ctx_per_seq = 4096
161
  llama_context: n_batch = 4096
162
  llama_context: n_ubatch = 4096
163
  llama_context: causal_attn = 1
164
+ llama_context: flash_attn = 1
165
  llama_context: kv_unified = true
166
  llama_context: freq_base = 1000000.0
167
  llama_context: freq_scale = 1
168
  llama_context: n_ctx_per_seq (4096) < n_ctx_train (128000) -- the full capacity of the model will not be utilized
169
+ llama_context: CUDA_Host output buffer size = 0.59 MiB
170
+ llama_kv_cache_unified: CUDA0 KV buffer size = 144.00 MiB
171
+ llama_kv_cache_unified: size = 144.00 MiB ( 4096 cells, 36 layers, 1/1 seqs), K (f16): 72.00 MiB, V (f16): 72.00 MiB
172
+ llama_context: CUDA0 compute buffer size = 2470.16 MiB
173
+ llama_context: CUDA_Host compute buffer size = 96.17 MiB
174
+ llama_context: graph nodes = 1234
175
+ llama_context: graph splits = 2
176
+ common_init_from_params: added <|endoftext|> logit bias = -inf
177
+ common_init_from_params: added <|im_end|> logit bias = -inf
178
+ common_init_from_params: added <|fim_pad|> logit bias = -inf
179
+ common_init_from_params: added <|repo_name|> logit bias = -inf
180
+ common_init_from_params: added <|file_sep|> logit bias = -inf
181
+ common_init_from_params: setting dry_penalty_last_n to ctx_size = 4096
182
+ common_init_from_params: warming up the model with an empty run - please wait ... (--no-warmup to disable)
183
 
184
  system_info: n_threads = 4 (n_threads_batch = 4) / 8 | CUDA : ARCHS = 890 | USE_GRAPHS = 1 | PEER_MAX_BATCH_SIZE = 128 | CPU : SSE3 = 1 | SSSE3 = 1 | AVX = 1 | AVX2 = 1 | F16C = 1 | FMA = 1 | BMI2 = 1 | AVX512 = 1 | AVX512_VNNI = 1 | LLAMAFILE = 1 | OPENMP = 1 | REPACK = 1 |
185
+ main: n_tokens in batch = 0
186
+ main: number of embeddings = 5090
187
  ```
188
 
189
+
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
190
 
191
  #### Table 2: NDCG@5
192
  ## NDCG@5 Performance Comparison