engineering/arborist
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

fix(#000057): real token usage + cost per input/output separately

Browse source 
fox 2026-05-21: (1) use REAL API token usage, not len//4; (2) the
substrate prefills a large retrieved CONTEXT as INPUT while solo feeds
~nothing, so per-completion-token over-charges the substrate — and per-
TOTAL-token UNDER-charges it (its mix is ~98% cheap prefill tokens).
Measured n=30 qwen-nothink/4090: substrate prefills ~6.6k input tok/query
(claim_lattice) vs solo ~52 — ~127x. Neither single per-token denominator
is honest; prefill (parallel, cheap/tok) and decode (autoregressive,
dear/tok) must be costed separately.

- OpenAICompatibleClient stashes data['usage'] as .last_usage (non-
  invasive; return type unchanged).
- watt_bench captures real prompt_tokens + completion_tokens per call
  (both arms), aggregates per cell, and energy_cogs reports gross +
  marginal per BOTH 1k-total-tok and 1k-completion-tok plus the context
  size. Prints the prompt/completion split.
- 12 tests incl. the prompt-context artifact (per-total cheap, per-
  completion dear). Full suite 2540 passed.

The clean per-input-tok / per-output-tok split rides bench/watt_calibrate
(slope calibration; separate commit once validated live).
This commit is contained in:
russell@unturf.com 2026-05-21 11:41:01 -04:00
parent 32aeb37086
commit 459060c774
No known key found for this signature in database
3 changed files with 131 additions and 68 deletions
Download patch file Download diff file Expand all files Collapse all files

156
bench/watt_bench.py
View file

@ -351,64 +351,82 @@ def _make_client(endpoint: str):
api_key=os.environ.get("ARBORIST_LLM_API_KEY"))
def _solo_call(client, mkey: str, question: str) -> tuple[str, int]:
"""One solo completion. Returns (answer, completion_tokens). The
bench client returns only the string; for token counts we re-issue
via a thin httpx call when usage is wanted — but to keep the energy
path simple we estimate tokens as len//4 when usage isn't surfaced."""
def _usage_tokens(client, answer: str) -> tuple[int, int]:
"""(prompt_tokens, completion_tokens) from the client's last call.
Real API usage when the endpoint reports it; falls back to a char//4
estimate for completion (prompt 0) only when usage is absent."""
u = getattr(client, "last_usage", None) or {}
p = u.get("prompt_tokens")
c = u.get("completion_tokens")
if c is None:
c = max(1, len(answer) // 4)
return int(p or 0), int(c)
def _solo_call(client, mkey: str, question: str) -> tuple[str, int, int]:
"""One solo completion. Returns (answer, prompt_tokens, completion_tokens)
from the endpoint's real usage. Solo's prompt is just system+question —
near-zero context — which is exactly the input asymmetry vs the
substrate arm (which prefills a large retrieved context)."""
cfg = MODELS[mkey]
ans = client.chat_completion(
[{"role": "system", "content": _SOLO_SYS},
{"role": "user", "content": question}],
model=cfg["model"], max_tokens=cfg["max_tokens"],
extra_body=cfg["extra"])
return ans, max(1, len(ans) // 4)
p, c = _usage_tokens(client, ans)
return ans, p, c
def energy_cogs(gpu_joules: float | None, window_s: float | None,
tot_tok: int, idle_mean_w: float | None,
completion_tok: int, idle_mean_w: float | None,
price_per_kwh: float,
serving_floor_w: float | None = None,
gen_draw_w: float | None = None,
gen_duty: float | None = None) -> dict:
"""Energy cost-of-goods-sold from MEASURED power states — no hardcodes.
gen_duty: float | None = None,
prompt_tok: int = 0) -> dict:
"""Energy cost-of-goods-sold from MEASURED power + REAL token usage.
fox 2026-05-21: a card occupies DISTINCT power states (idle, middle-
idle = model resident between requests, generation), and they differ
for every card × model × inference-server. So every power number is
measured at runtime and the states are kept DISTINCT — we do NOT
collapse them into ``joules/window`` and call that "generation"
(that blend is a duty-cycle artifact, not a state the card sits at).
Two corrections fox forced (2026-05-21):
* idle_mean_w — deep idle (no-request baseline window).
* serving_floor_w — middle-idle: model resident, between requests
(low band of the driven window). The right floor
for marginal cost — the standing cost of being
ready to answer.
* gen_draw_w — the actual generation draw (high band), with
``gen_duty`` = fraction of the window generating.
* marginal COGS — THE GPU COST OF A QUERY: joules ABOVE the
serving floor / tokens — the generation burst.
Retrieval/verify are CPU/SQLite (GPU untouched),
so this generation-only number is what the query
actually causes on the GPU. Falls back to idle
floor if no band split.
* gross — wall-clock window joules / tokens. NOT this
query's GPU cost: it charges the always-on
model-resident floor for the seconds we spend in
CPU-side retrieval (the card is free for other
traffic then). Kept for reference, not as COGS.
1. Power STATES, not a blend. A card occupies distinct states (idle,
middle-idle = model resident between requests, generation), which
differ per card×model×server. The marginal (energy above the
serving floor) is the query's GPU cost; ``gross`` (window integral)
is NOT query-attributable — it charges the always-on floor for the
seconds spent in CPU-side retrieval (GPU free for other traffic).
kWh = J / 3.6e6; ``$/1k-tok`` is the unit that compares to API pricing.
Only ``price_per_kwh`` is an operator input. See docs/stock-v1-config.md.
2. Cost per TOTAL tokens processed, not output only. The substrate
prefills a large retrieved CONTEXT (``prompt_tok``); solo prefills
almost nothing. An LLM request = prefill(all input) + decode(output),
so most of the substrate's GPU work is prefilling that context.
Dividing energy by completion tokens alone wrongly charges the
cost-of-reading-the-evidence to a few output tokens. We report BOTH
bases — per total (prompt+completion) is the apples-to-apples
compute cost; per completion is the answer-delivery view — and the
prompt-token count so the input asymmetry is explicit.
kWh = J / 3.6e6. ``price_per_kwh`` is the only operator input.
"""
if not gpu_joules or not tot_tok or not window_s:
completion_tok = completion_tok or 0
prompt_tok = prompt_tok or 0
total_tok = completion_tok + prompt_tok
if not gpu_joules or not window_s or total_tok <= 0:
return {"available": False, "price_per_kwh": price_per_kwh}
J_PER_KWH = 3.6e6
def per_1k(usd, tok):
return round(usd / tok * 1000, 6) if tok else None
gross_usd = gpu_joules / J_PER_KWH * price_per_kwh
out = {
"available": True,
"price_per_kwh": price_per_kwh,
# real token usage (the substrate's prompt_tok carries the context)
"prompt_tokens": prompt_tok,
"completion_tokens": completion_tok,
"total_tokens": total_tok,
# measured power states (distinct — not a blend)
"idle_w_measured": (round(idle_mean_w, 1)
if idle_mean_w is not None else None),
@ -421,7 +439,8 @@ def energy_cogs(gpu_joules: float | None, window_s: float | None,
"window_s": round(window_s, 1),
"gross_joules": round(gpu_joules, 1),
"gross_usd": round(gross_usd, 6),
"gross_usd_per_1k_tok": round(gross_usd / tot_tok * 1000, 6),
"gross_usd_per_1k_total_tok": per_1k(gross_usd, total_tok),
"gross_usd_per_1k_completion_tok": per_1k(gross_usd, completion_tok),
}
# Marginal against the serving floor (preferred) or idle (fallback).
floor_w = serving_floor_w if serving_floor_w is not None else idle_mean_w
@ -435,7 +454,9 @@ def energy_cogs(gpu_joules: float | None, window_s: float | None,
"marginal_floor_w": round(floor_w, 1),
"marginal_joules": round(marginal_joules, 1),
"marginal_usd": round(marginal_usd, 6),
"marginal_usd_per_1k_tok": round(marginal_usd / tot_tok * 1000, 6),
"marginal_usd_per_1k_total_tok": per_1k(marginal_usd, total_tok),
"marginal_usd_per_1k_completion_tok": per_1k(marginal_usd,
completion_tok),
})
return out
@ -552,7 +573,7 @@ def main() -> int:
tq = time.time()
cache_status = "solo_no_cache"
if arm == "solo":
ans, ctoks = _solo_call(client, mkey, q)
ans, ptoks, ctoks = _solo_call(client, mkey, q)
else: # arborist
from arborist.qa.query import query
from bench.stock_v1 import (
@ -583,7 +604,10 @@ def main() -> int:
extra_body=MODELS[mkey]["extra"],
policy=pol, burn_existing=True)
ans = r.get("raw_answer") or r.get("answer_text") or ""
ctoks = max(1, len(ans) // 4)
# Real usage from query()'s internal synthesis call
# (one call — STOCK V.1 repair is off). prompt_tokens
# carries the retrieved CONTEXT the substrate prefills.
ptoks, ctoks = _usage_tokens(client, ans)
cache_status = r.get("status", "?")
if (cache_status == "cache_hit"
or "cache_hit" in str(r.get("lookup_path", ""))):
@ -593,7 +617,8 @@ def main() -> int:
"variant": variant,
"latency_s": round(dt, 2),
"answer_chars": len(ans),
"est_completion_tokens": ctoks,
"prompt_tokens": ptoks,
"completion_tokens": ctoks,
"cache_status": cache_status})
finally:
client.close()
@ -603,7 +628,11 @@ def main() -> int:
bands = probe.band_stats()
elapsed = time.time() - t0
nq = len(rows)
tot_tok = sum(r["est_completion_tokens"] for r in rows)
tot_completion_tok = sum(r.get("completion_tokens",
r.get("est_completion_tokens", 0))
for r in rows)
tot_prompt_tok = sum(r.get("prompt_tokens", 0) for r in rows)
tot_tok = tot_completion_tok # back-compat alias for joules/token
gpu_j = st.get("joules")
cpu_j = cpu_st.get("joules")
total_j = (gpu_j or 0) + (cpu_j or 0) if (gpu_j or cpu_j) else None
@ -620,9 +649,11 @@ def main() -> int:
if cpu_j and nq else None),
"total_joules_per_question": (round(total_j / nq, 1)
if total_j and nq else None),
"est_total_completion_tokens": tot_tok,
"gpu_joules_per_token": (round(gpu_j / tot_tok, 3)
if gpu_j and tot_tok else None),
"total_completion_tokens": tot_completion_tok,
"total_prompt_tokens": tot_prompt_tok,
"total_tokens_processed": tot_completion_tok + tot_prompt_tok,
"gpu_joules_per_completion_token": (round(gpu_j / tot_completion_tok, 3)
if gpu_j and tot_completion_tok else None),
"mean_latency_s": (round(sum(r["latency_s"] for r in rows) / nq, 2)
if nq else None),
# Measured power-state decomposition (idle / serving-floor /
@ -639,12 +670,13 @@ def main() -> int:
# input. Window timestamps let a post-hoc load_monitor cross-ref
# flag organic-traffic contamination under non-isolation.
"energy_cogs": energy_cogs(
gpu_j, st.get("window_s"), tot_tok,
gpu_j, st.get("window_s"), tot_completion_tok,
(idle or {}).get("mean_w"), a.price_per_kwh,
serving_floor_w=(bands.get("low_band_w")
if bands.get("bimodal") else None),
gen_draw_w=bands.get("high_band_w"),
gen_duty=bands.get("duty_cycle")),
gen_duty=bands.get("duty_cycle"),
prompt_tok=tot_prompt_tok),
"window_start_unix": round(t0, 3),
"window_end_unix": round(t0 + elapsed, 3),
"watt_samples": (probe.watt_samples()
@ -673,17 +705,25 @@ def main() -> int:
+ ("" if b.get("bimodal") else " [unimodal — no gen state]"))
cg = cell["energy_cogs"]
if cg.get("available"):
# GPU COGS = generation energy ABOVE the always-on model-
# resident floor (the marginal). Retrieval/verify are CPU/
# SQLite — the GPU is untouched then, so the gross window
# integral OVER-counts (it charges the shared floor during our
# retrieval gap) and is NOT this query's GPU cost (fox).
print(f" GPU COGS @${cg['price_per_kwh']}/kWh: "
+ (f"${cg['marginal_usd_per_1k_tok']}/1k-tok generation "
f"(above {cg['marginal_floor']}-floor {cg['marginal_floor_w']}W)"
if "marginal_usd_per_1k_tok" in cg else "n/a")
+ f" · [window-total ${cg['gross_usd_per_1k_tok']}/1k-tok "
f"incl. shared idle during retrieval — NOT query-attributable]")
# Cost per TOTAL tokens processed (prompt+completion) is the
# apples-to-apples compute cost: the substrate prefills a large
# retrieved CONTEXT (prompt_tokens) that solo lacks, so per
# completion-token alone over-charges the substrate. GPU COGS =
# generation energy above the model-resident floor (retrieval/
# verify are CPU; GPU untouched). At our scale there is no
# prefill/KV-cache reuse (no policy yet for what's worth keeping
# hot), so that context is prefilled fresh every query.
print(f" ctx: prompt={cg['prompt_tokens']}tok "
f"completion={cg['completion_tokens']}tok "
f"(total {cg['total_tokens']})")
if "marginal_usd_per_1k_total_tok" in cg:
print(f" GPU COGS @${cg['price_per_kwh']}/kWh: "
f"${cg['marginal_usd_per_1k_total_tok']}/1k-total-tok "
f"· ${cg['marginal_usd_per_1k_completion_tok']}/1k-compl-tok "
f"(gen above {cg['marginal_floor']}-floor "
f"{cg['marginal_floor_w']}W)")
print(f" [window-total (ref, not attributable): "
f"${cg['gross_usd_per_1k_total_tok']}/1k-total-tok]")
for mkey in models:
run_cell("solo", mkey)