jartine commited on
Commit
0d16a68
1 Parent(s): 1550162

Update README.md

Browse files
Files changed (1) hide show
  1. README.md +8 -25
README.md CHANGED
@@ -26,10 +26,10 @@ NetBSD for AMD64 and ARM64.
26
 
27
  ## Quickstart
28
 
29
- Running the following on a desktop OS will launch a tab in your web
30
- browser with a chatbot interface.
31
 
32
  ```
 
33
  chmod +x llm-compiler-13b-ftd.Q6_K.llamafile
34
  ./llm-compiler-13b-ftd.Q6_K.llamafile --help
35
  ```
@@ -152,8 +152,7 @@ of consuming significantly less memory.
152
 
153
  ### Q6\_K, F16, BF16
154
 
155
- The output for the first xdigit() assembly optimization example with
156
- Q6\_K, F16, or BF16 weights is:
157
 
158
  ```asm
159
  .globl hiho
@@ -172,16 +171,10 @@ hiho: movl %edi, %ecx
172
  retq
173
  ```
174
 
175
- This implementation is **CORRECT**. This implementation is **GOOD**.
176
-
177
- NOTE: BF16 is currently only supported on CPU. It's the best quant for
178
- prompt processing on Zen4.
179
 
180
  ### Q5\_0
181
 
182
- The output for the first xdigit() assembly optimization example with
183
- Q5\_0 weights is:
184
-
185
  ```asm
186
  .globl hiho
187
  hiho: movl %edi, %ecx
@@ -204,15 +197,10 @@ hiho: movl %edi, %ecx
204
  retq
205
  ```
206
 
207
- This implementation is **INCORRECT**. Even if it did work, it has
208
- unneeded stack spillage and branches. Therefore this quantization format
209
- was rejected and isn't included in this repository.
210
 
211
  ### Q5\_K\_M
212
 
213
- The output for the first xdigit() assembly optimization example with
214
- Q5\_K\_M weights is:
215
-
216
  ```asm
217
  .globl hiho
218
  hiho: movl $48, %eax
@@ -245,14 +233,11 @@ hiho: movl $48, %eax
245
  retq
246
  ```
247
 
248
- This implementation is **INCORRECT**. Even if it did work, it has
249
- unneeded stack spillage and branches. Therefore this quantization format
250
- was rejected and isn't included in this repository.
251
 
252
  ### Q8\_0
253
 
254
- The output for the first xdigit() assembly optimization example with
255
- Q8\_0 weights is:
256
 
257
  ```asm
258
  .globl hiho
@@ -286,9 +271,7 @@ hiho: movl $48, %eax
286
  retq
287
  ```
288
 
289
- This implementation is **INCORRECT**. Even if it did work, it has
290
- unneeded stack spillage and branches. Therefore this quantization format
291
- was rejected and isn't included in this repository.
292
 
293
  ---
294
 
 
26
 
27
  ## Quickstart
28
 
29
+ Running the following will display the help:
 
30
 
31
  ```
32
+ wget https://huggingface.co/Mozilla/llm-compiler-13b-ftd-llamafile/resolve/main/llm-compiler-13b-ftd.Q6_K.llamafile
33
  chmod +x llm-compiler-13b-ftd.Q6_K.llamafile
34
  ./llm-compiler-13b-ftd.Q6_K.llamafile --help
35
  ```
 
152
 
153
  ### Q6\_K, F16, BF16
154
 
155
+ The output for optimizing xdigit() is:
 
156
 
157
  ```asm
158
  .globl hiho
 
171
  retq
172
  ```
173
 
174
+ This implementation is **CORRECT**.
 
 
 
175
 
176
  ### Q5\_0
177
 
 
 
 
178
  ```asm
179
  .globl hiho
180
  hiho: movl %edi, %ecx
 
197
  retq
198
  ```
199
 
200
+ This implementation is **INCORRECT**.
 
 
201
 
202
  ### Q5\_K\_M
203
 
 
 
 
204
  ```asm
205
  .globl hiho
206
  hiho: movl $48, %eax
 
233
  retq
234
  ```
235
 
236
+ This implementation is **INCORRECT**.
 
 
237
 
238
  ### Q8\_0
239
 
240
+ The output for optimizing xdigit() is:
 
241
 
242
  ```asm
243
  .globl hiho
 
271
  retq
272
  ```
273
 
274
+ This implementation is **INCORRECT**.
 
 
275
 
276
  ---
277