Hugging Face's logo

Spaces:
canada-quant
/
README
Running

App Files Community
pastapaul commited on
Commit
1d88966
·
verified ·
1 Parent(s): 3cf8366

Replace default static-space welcome with proper org card (lab intro + DeepSeek-V4 family table)

Browse files
Files changed (1) hide show
  1. index.html +255 -17
index.html CHANGED
@@ -1,19 +1,257 @@
1
  <!doctype html>
2
- <html>
3
- <head>
4
- <meta charset="utf-8" />
5
- <meta name="viewport" content="width=device-width" />
6
- <title>My static Space</title>
7
- <link rel="stylesheet" href="style.css" />
8
- </head>
9
- <body>
10
- <div class="card">
11
- <h1>Welcome to your static Space!</h1>
12
- <p>You can modify this app directly by editing <i>index.html</i> in the Files and versions tab.</p>
13
- <p>
14
- Also don't forget to check the
15
- <a href="https://huggingface.co/docs/hub/spaces" target="_blank">Spaces documentation</a>.
16
- </p>
17
- </div>
18
- </body>
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
19
  </html>
 
1
  <!doctype html>
2
+ <html lang="en">
3
+ <head>
4
+ <meta charset="utf-8" />
5
+ <meta name="viewport" content="width=device-width, initial-scale=1" />
6
+ <title>Canada Quant Labs</title>
7
+ <style>
8
+ :root {
9
+ --red: #d52b1e;
10
+ --ink: #111;
11
+ --muted: #5a5a5a;
12
+ --rule: #e5e5e5;
13
+ --bg: #ffffff;
14
+ --code-bg: #f6f6f6;
15
+ }
16
+ * { box-sizing: border-box; }
17
+ html, body {
18
+ margin: 0;
19
+ padding: 0;
20
+ background: var(--bg);
21
+ color: var(--ink);
22
+ font-family: -apple-system, BlinkMacSystemFont, "Segoe UI", Inter, Roboto, "Helvetica Neue", Arial, sans-serif;
23
+ line-height: 1.55;
24
+ font-size: 15px;
25
+ }
26
+ .wrap {
27
+ max-width: 980px;
28
+ margin: 0 auto;
29
+ padding: 32px 24px 48px;
30
+ }
31
+ header {
32
+ border-bottom: 3px solid var(--red);
33
+ padding-bottom: 18px;
34
+ margin-bottom: 24px;
35
+ }
36
+ h1 {
37
+ font-size: 32px;
38
+ font-weight: 700;
39
+ margin: 0 0 6px;
40
+ letter-spacing: -0.01em;
41
+ }
42
+ h1 .leaf { color: var(--red); }
43
+ .tagline {
44
+ font-size: 16px;
45
+ color: var(--muted);
46
+ margin: 0;
47
+ }
48
+ h2 {
49
+ font-size: 20px;
50
+ font-weight: 600;
51
+ margin: 32px 0 12px;
52
+ letter-spacing: -0.005em;
53
+ }
54
+ h3 {
55
+ font-size: 15px;
56
+ font-weight: 600;
57
+ margin: 22px 0 8px;
58
+ color: var(--muted);
59
+ text-transform: uppercase;
60
+ letter-spacing: 0.04em;
61
+ }
62
+ p { margin: 0 0 12px; }
63
+ ul {
64
+ margin: 0 0 14px;
65
+ padding-left: 20px;
66
+ }
67
+ li { margin-bottom: 4px; }
68
+ a {
69
+ color: var(--red);
70
+ text-decoration: none;
71
+ border-bottom: 1px solid transparent;
72
+ transition: border-color 0.15s;
73
+ }
74
+ a:hover { border-bottom-color: var(--red); }
75
+ code {
76
+ background: var(--code-bg);
77
+ padding: 1px 5px;
78
+ border-radius: 3px;
79
+ font-family: ui-monospace, "SF Mono", Menlo, Consolas, monospace;
80
+ font-size: 0.9em;
81
+ }
82
+ .pillars {
83
+ display: grid;
84
+ grid-template-columns: repeat(auto-fit, minmax(220px, 1fr));
85
+ gap: 18px;
86
+ margin: 18px 0 8px;
87
+ }
88
+ .pillar {
89
+ border-left: 3px solid var(--red);
90
+ padding: 4px 0 4px 14px;
91
+ }
92
+ .pillar-title {
93
+ font-weight: 600;
94
+ font-size: 13px;
95
+ color: var(--muted);
96
+ text-transform: uppercase;
97
+ letter-spacing: 0.04em;
98
+ margin-bottom: 6px;
99
+ }
100
+ .pillar-body { font-size: 14px; }
101
+ .table-scroll {
102
+ overflow-x: auto;
103
+ margin: 8px 0 16px;
104
+ border: 1px solid var(--rule);
105
+ border-radius: 6px;
106
+ }
107
+ table {
108
+ width: 100%;
109
+ border-collapse: collapse;
110
+ font-size: 13.5px;
111
+ }
112
+ th, td {
113
+ text-align: left;
114
+ padding: 10px 12px;
115
+ border-bottom: 1px solid var(--rule);
116
+ vertical-align: top;
117
+ }
118
+ th {
119
+ background: #fafafa;
120
+ font-weight: 600;
121
+ white-space: nowrap;
122
+ }
123
+ tr:last-child td { border-bottom: none; }
124
+ td code { font-size: 12.5px; }
125
+ .meta-line {
126
+ font-size: 13.5px;
127
+ color: var(--muted);
128
+ margin: 14px 0;
129
+ }
130
+ .meta-line strong { color: var(--ink); }
131
+ .contact {
132
+ margin-top: 28px;
133
+ padding-top: 18px;
134
+ border-top: 1px solid var(--rule);
135
+ font-size: 14px;
136
+ color: var(--muted);
137
+ }
138
+ .contact a { color: var(--ink); }
139
+ .contact a:hover { color: var(--red); border-bottom-color: var(--red); }
140
+ .repos-list {
141
+ font-size: 13.5px;
142
+ list-style: none;
143
+ padding-left: 0;
144
+ margin-top: 4px;
145
+ }
146
+ .repos-list li { margin-bottom: 3px; }
147
+ </style>
148
+ </head>
149
+ <body>
150
+ <div class="wrap">
151
+
152
+ <header>
153
+ <h1><span class="leaf">🍁</span> Canada Quant Labs</h1>
154
+ <p class="tagline">Canada's open-weight model lab. Sovereign AI on Canadian Blackwell silicon for regulated industries.</p>
155
+ </header>
156
+
157
+ <div class="pillars">
158
+ <div class="pillar">
159
+ <div class="pillar-title">What we do</div>
160
+ <div class="pillar-body">Post-training on open base models (SFT, DPO, GRPO, RLAIF) · Production quantization recipes (W4A16, NVFP4, MXFP4) · Audited, air-gapped deployment with eval evidence and MRM docs.</div>
161
+ </div>
162
+ <div class="pillar">
163
+ <div class="pillar-title">Where we work</div>
164
+ <div class="pillar-body">Legal · Medical · Defence · Finance. Headquarters in Victoria, BC. Compute on NVIDIA DGX B300 at Equinix Vancouver.</div>
165
+ </div>
166
+ <div class="pillar">
167
+ <div class="pillar-title">Upstream</div>
168
+ <div class="pillar-body">Contributors to vLLM, llm-compressor, compressed-tensors. PR <a href="https://github.com/vllm-project/vllm/pull/42209">#42209</a> (NVFP4 MoE for DSV4) merged 2026-05-22 with our validation.</div>
169
+ </div>
170
+ </div>
171
+
172
+ <h2>Open releases — DeepSeek-V4 quantization family</h2>
173
+
174
+ <p>Four artifacts in the same lineage. One base model in two sizes (V4-Flash, V4-Pro); two routed-expert formats (W4A16, NVFP4); Multi-Token Prediction (MTP) draft head retained on three of four. Attention is FP8 block 128×128 across all four.</p>
175
+
176
+ <div class="table-scroll">
177
+ <table>
178
+ <thead>
179
+ <tr>
180
+ <th>Model</th>
181
+ <th>Base</th>
182
+ <th>Routed experts</th>
183
+ <th>MTP</th>
184
+ <th>On-disk</th>
185
+ <th>Min hardware (TP=2)</th>
186
+ <th>When to pick</th>
187
+ </tr>
188
+ </thead>
189
+ <tbody>
190
+ <tr>
191
+ <td><a href="https://huggingface.co/canada-quant/DeepSeek-V4-Flash-W4A16-FP8">V4-Flash-W4A16-FP8</a></td>
192
+ <td>V4-Flash</td>
193
+ <td>W4A16 INT4 g=128</td>
194
+ <td>no</td>
195
+ <td>~143 GB</td>
196
+ <td>H200 / DGX Spark / RTX PRO 6000</td>
197
+ <td>maximum compatibility, no MTP needed</td>
198
+ </tr>
199
+ <tr>
200
+ <td><a href="https://huggingface.co/canada-quant/DeepSeek-V4-Flash-W4A16-FP8-MTP">V4-Flash-W4A16-FP8-MTP</a></td>
201
+ <td>V4-Flash</td>
202
+ <td>W4A16 INT4 g=128</td>
203
+ <td>yes (BF16)</td>
204
+ <td>159 GB</td>
205
+ <td>H200 / RTX PRO 6000</td>
206
+ <td>best $/token interactive on V4-Flash</td>
207
+ </tr>
208
+ <tr>
209
+ <td><a href="https://huggingface.co/canada-quant/DeepSeek-V4-Flash-NVFP4-FP8-MTP">V4-Flash-NVFP4-FP8-MTP</a></td>
210
+ <td>V4-Flash</td>
211
+ <td>NVFP4 g=16</td>
212
+ <td>yes (BF16)</td>
213
+ <td>172 GB</td>
214
+ <td>RTX PRO 6000 / B300</td>
215
+ <td>best Blackwell-native interactive on V4-Flash</td>
216
+ </tr>
217
+ <tr>
218
+ <td><a href="https://huggingface.co/canada-quant/DeepSeek-V4-Pro-NVFP4-FP8-MTP">V4-Pro-NVFP4-FP8-MTP</a></td>
219
+ <td>V4-Pro</td>
220
+ <td>NVFP4 g=16</td>
221
+ <td>yes (byte-identical)</td>
222
+ <td>913 GiB</td>
223
+ <td>8× B300 (TP=8 + EP)</td>
224
+ <td>only choice for V4-Pro; <strong>+25–37% throughput vs upstream MXFP4</strong></td>
225
+ </tr>
226
+ </tbody>
227
+ </table>
228
+ </div>
229
+
230
+ <p class="meta-line">Upstream reference recipes: <a href="https://huggingface.co/RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8"><code>RedHatAI/DeepSeek-V4-Flash-NVFP4-FP8</code></a> (Flash NVFP4 topology) and <a href="https://huggingface.co/nvidia/DeepSeek-V3.2-NVFP4"><code>nvidia/DeepSeek-V3.2-NVFP4</code></a> (Pro NVFP4, MTP-exclusion topology).</p>
231
+
232
+ <h3>Hardware shorthand</h3>
233
+ <ul>
234
+ <li><strong>H200</strong> — 8× NVIDIA H200 SXM5 (Hopper SM 9.0a, 141 GB HBM3e/GPU)</li>
235
+ <li><strong>DGX Spark</strong> — 2× NVIDIA DGX Spark (GB10, Blackwell SM 12.1a)</li>
236
+ <li><strong>RTX PRO 6000</strong> — NVIDIA RTX PRO 6000 Blackwell Server Edition (SM 12.0, sm_120, 96 GB HBM)</li>
237
+ <li><strong>B300</strong> — NVIDIA B300 SXM6 AC (Blackwell SM 10.3, sm_103a, 288 GB HBM3e/GPU)</li>
238
+ </ul>
239
+
240
+ <h3>Reproduction repos</h3>
241
+ <p>Every artifact has a public reproduction repo with calibration scripts, vLLM patches, bench harnesses, and findings docs:</p>
242
+ <ul class="repos-list">
243
+ <li>↳ <a href="https://github.com/canada-quant/dsv4-flash-w4a16-fp8"><code>canada-quant/dsv4-flash-w4a16-fp8</code></a></li>
244
+ <li>↳ <a href="https://github.com/canada-quant/dsv4-flash-w4a16-fp8-mtp"><code>canada-quant/dsv4-flash-w4a16-fp8-mtp</code></a></li>
245
+ <li>↳ <a href="https://github.com/canada-quant/dsv4-flash-nvfp4-fp8-mtp"><code>canada-quant/dsv4-flash-nvfp4-fp8-mtp</code></a></li>
246
+ <li>↳ <a href="https://github.com/canada-quant/dsv4-pro-nvfp4-fp8-mtp"><code>canada-quant/dsv4-pro-nvfp4-fp8-mtp</code></a></li>
247
+ </ul>
248
+
249
+ <div class="contact">
250
+ Partnerships · <a href="mailto:partnerships@cql.ca">partnerships@cql.ca</a> &nbsp;&middot;&nbsp;
251
+ Press · <a href="mailto:press@cql.ca">press@cql.ca</a> &nbsp;&middot;&nbsp;
252
+ Web · <a href="https://cql.ca">cql.ca</a>
253
+ </div>
254
+
255
+ </div>
256
+ </body>
257
  </html>