jgalego commited on
Commit
916de58
·
verified ·
1 Parent(s): 6b87b2c

Upload folder using huggingface_hub

Browse files
README.md CHANGED
@@ -33,20 +33,7 @@ It ports [Qwenputer](https://github.com/divagr18/qwenputer) (the same machine in
33
 
34
  ### A calc call
35
 
36
- ```mermaid
37
- sequenceDiagram
38
- participant U as User
39
- participant O as OLMo's 32 layers
40
- participant C as The chip layers (the CPU)
41
- U->>O: What is 123456789 * 987654321?
42
- O->>C: function_calls opens: calc(expression='123456789*987654321')
43
- Note over C: every token of the call is copied into the text buffer the firmware reads
44
- O->>C: function_calls closes
45
- Note over C: the CPU runs the call: 2,717 trace tokens through the chip layers only,<br/>which OLMo's layers, caches and positions never see
46
- C->>O: im_end, im_start, environment, 121 932 631 112 635 269, im_end, im_start, assistant
47
- Note over O: each echoed token is text: it runs through every layer
48
- O->>U: The result of 123456789 multiplied by 987654321 is 121932631112635269.
49
- ```
50
 
51
 
52
 
@@ -54,18 +41,7 @@ sequenceDiagram
54
 
55
  Each trace token runs once through the 11 chip layers, and each layer does one part of executing the instruction:
56
 
57
- ```mermaid
58
- flowchart TB
59
- IN["a trace token in"] --> L1["L1: the PC anchor and counters (record position, commits since the last jump)"]
60
- L1 --> L2["L2: the PC word and the link address"]
61
- L2 --> L3["L3: fetch: the code ROM, 23 heads of 2,944 words each"]
62
- L3 --> L4["L4: decode; read 3 register ports"]
63
- L4 --> L5["L5: the bignum digit file; adder and multiplier stages"]
64
- L5 --> L6["L6 to L8: the RAM, 86 heads of 96 words; loads, stores, compares"]
65
- L6 --> L9["L9: the data ROM; divider digits"]
66
- L9 --> L10["L10 to L11: pick the result and write the next token's code"]
67
- L10 --> OUT["the next trace token"]
68
- ```
69
 
70
  ![Heads per chip layer, by the memory they hold](assets/layers.png)
71
 
 
33
 
34
  ### A calc call
35
 
36
+ ![One calc call: who writes each part of the stream](assets/diagram-call.png)
 
 
 
 
 
 
 
 
 
 
 
 
 
37
 
38
 
39
 
 
41
 
42
  Each trace token runs once through the 11 chip layers, and each layer does one part of executing the instruction:
43
 
44
+ ![What each chip layer computes for one trace token](assets/diagram-layers.png)
 
 
 
 
 
 
 
 
 
 
 
45
 
46
  ![Heads per chip layer, by the memory they hold](assets/layers.png)
47
 
assets/diagram-call.png ADDED
assets/diagram-layers.png ADDED