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

feat(us-026): smart model assignment via demand×coverage scoring

Browse Source 
/v1/network/assign now scores models by (demand_rpm + 1) × (coverage_deficit + 0.01)
so high-traffic, under-covered models are preferred when assigning new nodes.
Response includes price_per_token: 0.0 (reserved for future pricing protocol).
--memory MB flag added to node CLI to override autodetected VRAM budget for
shard assignment without changing hardware detection for inference.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Dobromir Popov
2026-06-30 13:42:43 +03:00
parent d9110b623b
commit 27818df654
3 changed files with 33 additions and 6 deletions
Download Patch File Download Diff File Expand all files Collapse all files

25
packages/tracker/meshnet_tracker/server.py
View File

@@ -1550,8 +1550,17 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
if n.hf_repo not in repo_layers or n.num_layers > repo_layers[n.hf_repo]:
repo_layers[n.hf_repo] = n.num_layers
# Pick the repo where the gap is largest (most work to do).
# Smart scoring: demand_rpm × coverage_deficit
# coverage_deficit = uncovered_layers / total_layers (0 = fully covered)
# demand_rpm comes from the stats collector; defaults to 0.0 when no traffic yet.
demand_rpms: dict[str, float] = {}
if server.stats is not None:
local_rpms = server.stats.get_local_rpms()
for repo in repo_groups:
demand_rpms[repo] = local_rpms.get(repo, {}).get("rpm_last_hour", 0.0)
best_repo = None
best_score = -1.0
best_gap_size = -1
best_gap_start = 0
best_num_layers = 0
@@ -1569,8 +1578,13 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
gap_start = max(gap_start, e + 1)
else:
break
gap_size = max(0, (total - 1) - gap_start + 1) # layers remaining uncovered
if gap_size > best_gap_size:
gap_size = max(0, (total - 1) - gap_start + 1)
coverage_deficit = gap_size / max(total, 1)
demand = demand_rpms.get(repo, 0.0)
# +1.0 floor on demand so models with no recorded traffic still compete by coverage.
score = (demand + 1.0) * (coverage_deficit + 0.01)
if score > best_score or (score == best_score and gap_size > best_gap_size):
best_score = score
best_gap_size = gap_size
best_gap_start = gap_start
best_repo = repo
@@ -1578,8 +1592,8 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
gap_found = best_gap_size > 0
if not gap_found:
# All shards are covered — still assign to the model with most nodes for redundancy.
best_repo = max(repo_groups, key=lambda r: len(repo_groups[r]))
# All shards covered — assign to highest-demand model for redundancy.
best_repo = max(repo_groups, key=lambda r: demand_rpms.get(r, 0.0))
best_gap_start = 0
best_num_layers = repo_layers[best_repo]
@@ -1599,6 +1613,7 @@ class _TrackerHandler(http.server.BaseHTTPRequestHandler):
"shard_end": shard_end,
"num_layers": total_l,
"gap_found": gap_found,
"price_per_token": 0.0,
})
def _handle_route(self, parsed: urllib.parse.ParseResult):