xinfer-ai 0.13.2 → 0.13.3
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +11 -0
- package/docs/performance.md +1 -1
- package/package.json +2 -2
- package/scripts/install.js +4 -1
package/README.md
CHANGED
|
@@ -460,6 +460,17 @@ Constraint-based generation via llguidance — Lark grammars, regex, JSON Schema
|
|
|
460
460
|
| `--mcp-config` | MCP servers JSON config |
|
|
461
461
|
| `--mcp-command` / `--mcp-args` | Single MCP server command + args |
|
|
462
462
|
|
|
463
|
+
### Environment Variables
|
|
464
|
+
|
|
465
|
+
| Variable | Description |
|
|
466
|
+
|---|---|
|
|
467
|
+
| `XINFER_NVFP4_FORCE_LUT=1` | Force software NVFP4 decode to use the LUT-based dequantization path (higher precision) instead of hardware FP4 intrinsics on Blackwell (SM100+). Useful when decode precision matters more than peak throughput. |
|
|
468
|
+
|
|
469
|
+
**Example (Blackwell with high-precision NVFP4 decode):**
|
|
470
|
+
```bash
|
|
471
|
+
XINFER_NVFP4_FORCE_LUT=1 xinfer --m nvidia/Qwen3-30B-A3B-FP4 --ui-server
|
|
472
|
+
```
|
|
473
|
+
|
|
463
474
|
---
|
|
464
475
|
|
|
465
476
|
## 📽️ Demo
|
package/docs/performance.md
CHANGED
|
@@ -50,7 +50,7 @@ xinfer --m AxionML/Qwen3.5-4B-NVFP4 --kvcache-dtype turbo4 --ui-server
|
|
|
50
50
|
|
|
51
51
|
## 📊 Performance Notes
|
|
52
52
|
|
|
53
|
-
- **HW NVFP4**: Hardware-accelerated FP4 on Blackwell (
|
|
53
|
+
- **HW NVFP4**: Hardware-accelerated FP4 on Blackwell: SM100 (B100/B200), SM120 (RTX 5090/RTX PRO 6000), and SM121 (DGX Spark)
|
|
54
54
|
- **HW FP8**: Hardware-accelerated FP8 on Hopper (SM90, H800/H200)
|
|
55
55
|
- **SW FP4/NVFP4**: Software-emulated FP4 on Hopper (SM90) and V100 (SM70)
|
|
56
56
|
- V100 + NVFP4 + TurboQuant is a first-ever combination — no other engine has achieved this
|
package/package.json
CHANGED
|
@@ -1,7 +1,7 @@
|
|
|
1
1
|
{
|
|
2
2
|
"name": "xinfer-ai",
|
|
3
|
-
"version": "0.13.
|
|
4
|
-
"assetVersion": "0.13.
|
|
3
|
+
"version": "0.13.3",
|
|
4
|
+
"assetVersion": "0.13.3",
|
|
5
5
|
"description": "xInfer — a high-performance LLM inference engine in Rust with CUDA/Metal acceleration",
|
|
6
6
|
"license": "MIT",
|
|
7
7
|
"homepage": "https://github.com/guoqingbao/xinfer#readme",
|
package/scripts/install.js
CHANGED
|
@@ -75,11 +75,14 @@ function detectCudaComputeCap() {
|
|
|
75
75
|
|
|
76
76
|
/**
|
|
77
77
|
* Map detected SM version to the build variant name.
|
|
78
|
-
* Build variants match CI matrix: sm70, sm75, sm80, sm89, sm90,
|
|
78
|
+
* Build variants match CI matrix: sm70, sm75, sm80, sm89, sm90, sm100,
|
|
79
|
+
* sm120, and sm121 (GB10/DGX Spark).
|
|
79
80
|
*/
|
|
80
81
|
function smToVariant(sm) {
|
|
81
82
|
const n = parseInt(sm, 10);
|
|
83
|
+
if (n >= 121) return "sm121";
|
|
82
84
|
if (n >= 120) return "sm120";
|
|
85
|
+
if (n >= 100) return "sm100";
|
|
83
86
|
if (n >= 90) return "sm90";
|
|
84
87
|
if (n >= 89) return "sm89";
|
|
85
88
|
if (n >= 80) return "sm80";
|