Endless loop after jax.jacrev(ode solver)(args)

Hello,
I have a simple (I think so) use-case with

def ode(t, M, args):
    x,y,z,w = M
    a,b,c,d,lam,mu,g,h,f,p,q = args
    #
    ....
    #
    return jnp.array([d_x, d_y, d_z, d_w])


t0 = 0.
t1 = 10000
dt0 = 1e-2

# init value
x0 = 0.01
y0 = 1.0
z0 = 0.01
w0 = 1.0
M0 = jnp.array([x0,y0,z0,w0])


@jax.jit
def run(params):
  # params = a,b,c,d,lam,mu,g,h,f,p,q 

  term = ODETerm(ode)
  solver = Dopri5()

  saveat = SaveAt(ts=jnp.linspace(t0, t1, 10000))

  sols = diffeqsolve(terms=term, solver=solver, t0=t0, t1=t1,dt0=dt0,
                       y0=M0, args=params,
                       saveat=saveat, max_steps=100_000, throw=False)

  return sols.ys[:,0] # x

# to get x(t) : that's ok 
#tmp= run(params)

# to get dx/dparams  THIS IS THE QUESTION
tmp= jax.jacrev(run)(params)

The diemnsion of tmp is (10000,11) ( number of time saved, number of params)

The problem is that something a simple as

tmp[:10,0]

to get the first 10 values of dx/a (a: first parameter) is starting a endless loop.
Any idea, where is my mistake?
Thanks

Reproduce the shown example, comparing run(params) with jax.jacrev(run)(params) and then the tmp[:10,0] access. Inspect the interaction between jax.jacrev, the @jax.jit-wrapped run function, and diffeqsolve with Dopri5 and SaveAt; done means the Jacobian can be sliced without an endless loop.