popov/neuron-tai
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

5-th DL fix

Browse Source
This commit is contained in:
Dobromir Popov
2026-07-06 22:55:01 +03:00
parent 7e7682be47
commit 4bfdc814e2
6 changed files with 126 additions and 78 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
packages/tracker/meshnet_tracker/server.py
View File

@@ -3758,7 +3758,9 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
except ValueError:
ram_mb = 0
max_layers = required_end - required_start + 1
memory_mb = vram_mb if vram_mb > 0 else ram_mb
# VRAM only bounds the shard for CUDA nodes; a CPU node may still report
# a detected-but-unusable GPU, and must be sized by system RAM.
memory_mb = vram_mb if (device == "cuda" and vram_mb > 0) else ram_mb
if memory_mb > 0:
layer_bytes = _preset_bytes_per_layer(preset).get("bfloat16", 30 * 1024 * 1024)
max_layers = min(max_layers, max(1, int(((memory_mb * 1024 * 1024) * 0.8) // layer_bytes)))
@@ -3963,7 +3965,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
if preset is not None and preset.get("hf_repo"):
required_start, required_end = _preset_layer_bounds(preset)
total_l = required_end - required_start + 1
memory_mb = vram_mb if vram_mb > 0 else ram_mb
memory_mb = vram_mb if (device == "cuda" and vram_mb > 0) else ram_mb
max_layers = _max_layers_for_memory(memory_mb, total_l, preset)
shard_start = required_start
shard_end = min(required_end, shard_start + max_layers - 1)
@@ -4052,7 +4054,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
# Capacity: use the same 80%-of-memory rule as registered node planning.
total_l = best_num_layers
memory_mb = vram_mb if vram_mb > 0 else ram_mb
memory_mb = vram_mb if (device == "cuda" and vram_mb > 0) else ram_mb
resolved_name, best_preset = _resolve_model_preset(server.model_presets, str(best_repo))
if memory_mb > 0:
max_layers = _max_layers_for_memory(memory_mb, total_l, best_preset)